News Weekly
ENहिन्दी
LV 10 XP
0% read
Your progress · 0/5 chapters
About 20 min total
Software Engineering & SDLCISSUE #7 · STORY 5 OF 10Oct 16, 2026CONFIRMED

Replit makes OpenAI its default model over Anthropic

Replit declares OpenAI its preferred partner for Replit Agent while clarifying it's still model-agnostic, as Claude models remain available across paid tiers.

Schematic routing junction showing a preferred track alongside intact parallel tracks, representing platform preference coexisting with continued model availability.

Read it your way

CHAPTER 1Picked for Explorer

The 60-second version

At TechCrunch Disrupt, Replit's CEO said 'no more Anthropic' but quickly clarified OpenAI is preferred while remaining model-agnostic, as Claude models are still selectable and even added.

Headline vs realityInitial 'no longer working' was walked back to 'preferred partner, model agnostic'.
Claude still thereFive Claude models remain in Max tier, two in Power, with Haiku 5.5 added Oct 16.
Defaults matterPreference likely affects Free Mode and Auto routing, not manual selection.
Claims unverifiedPerformance/cost claims contradict Replit's own earlier ViBench results.
Finish this chapter for +15 XP
Abstract junction selector showing preference for one route while keeping all parallel routes selectable, representing routing defaults coexisting with model optionality.
Your next move · Explorer

Verify before believing headlines

1Check actual product features against news claims
2Learn to distinguish routing defaults from full migrations
3Track model availability across platforms over time

Switch your reading mode at the top to see a different next move.

Tap to open

Things to keep an eye on

Pop quiz · unlock the Model Wars badge

Did it stick?

0/3
According to the clarification, is Replit still working with Anthropic?+20 XP
Where would a 'default model' change have the most impact?+20 XP
What happened to Claude models on Replit after the Disrupt remarks?+20 XP
Your call · +5 XP

Should coding agent platforms remain model-agnostic?

Research dossierThe full research, labeled and sourcedCONFIRMED22 sources · 20 min read · 21 sections
Story identity
  • Story ID: S09
  • Title: Replit makes OpenAI its default model over Anthropic
  • Category: software-engineering (developer tooling / agentic coding; model-provider competition)
  • Evidence status: split, and the split is the story:
    • The statement event — CONFIRMED. Masad said it, on record, on stage at Disrupt; TechCrunch reported it with direct quotes; a session recording exists (https://www.youtube.com/watch?v=grkRhVKcxiE); the recap and relays are consistent.
    • The substance — COMPANY CLAIM. That OpenAI is now Replit's default / preferred model provider rests entirely on Replit's own statements (Masad on stage + his written clarification). A person familiar with the partnership told TechCrunch the Anthropic partnership is still in place, and live public artifacts show Anthropic models still shipping in Replit Agent (see §5). No independent audit of routing defaults exists.

Evidence labels used below: CONFIRMED, COMPANY CLAIM, INDEPENDENT EVIDENCE, EARLY RESEARCH, PRESS REPORT (relay), INTERPRETATION, PREDICTION.

Core claim split (critical)

ClaimWhat is actually evidencedLabel
OpenAI is Replit's preferred partner powering Replit Agent; Replit "remains model agnostic"Masad's written clarification of 2026-10-17, relayed verbatim by zeitgraphs and WebProNewsCOMPANY CLAIM (company statement, two relays; direct X post not retrieved)
·

Runtime context

  • Research window (from, sole source of truth): 2026-10-14 → 2026-10-17 (inclusive).
  • story_count: 10. This artifact covers exactly one assigned story: S09.
  • Window validation: actual event date 2026-10-15 (Amjad Masad's on-stage remarks at TechCrunch Disrupt 2026) falls inside the window — PASS. event_date of 2026-10-16 is the news-publication date, not the on-stage event date; it too falls inside the window, so window validation passes under either reading. The discrepancy is recorded below, not silently corrected.
  • Context events outside the window are explicitly labelled as context (Free Mode post 2026-08-18; OpenAI partnership page 2026-08-19; Replit Agent on Claude 3.5 Sonnet 2024-09; Inc. revenue story 2025-02-18; ViBench 2026-06-23; "Free the models" 2026-09-29; changelog entry 2026-10-02; funding profile 2026-09-30). The 2026-10-16 changelog addition (Claude Haiku 5.5) falls inside the window and is treated as evidence.

Date record (kept separate)

Date typeValueEvidence
Actual event date — Masad's remarks2026-10-15 (Thursday) — on stage at TechCrunch Disrupt 2026 (Moscone West, San Francisco), during the session on the "self-driving company", in reply to an interviewer question about whether the August OpenAI partnership meant "no more Anthropic"TechCrunch session recap (2026-10-16) says the remarks were made "on Thursday"; WebProNews dates them "Thursday, Oct. 15"; TechCrunch Disrupt agenda page shows the event ran 2026-10-13 → 2026-10-15; Unite AI relay of the exchange
Discovery-recorded event_date2026-10-16 — recorded by discovery; corresponds to the publication date of the TechCrunch article, not the on-stage event. Discrepancy recorded; not used as the event date hereS09 record vs. TechCrunch article timestamps and recap wording
Announcement date (news publication)2026-10-16, 2:22 pm PT — TechCrunch publishes "Replit is no longer working with Anthropic, CEO says" (original URL path /2026/10/16/…, still indexed)TechCrunch article ( 2026-10-18; re-titled/updated version carries the /2026/10/17/… path)
Article update / retitle2026-10-17, 11:48 AM PDT — updated, headline softened to reflect Masad's written clarification; URL moved to /2026/10/17/replit-is-no-longer-working-with-anthropic-ceo-says/TechCrunch article byline block
Company clarification date2026-10-17 — Masad posts a written clarification ("Replit remains model agnostic… OpenAI is now our preferred partner…") on X; the direct X status URL was not retrieved (see §13)Full text relayed by zeitgraphs and WebProNews
Article dates2026-10-16: TechCrunch main piece (original), TechCrunch session recap; 2026-10-17: TechCrunch updated piece, WebProNews relay of the clarification; Unite AI, zeitgraphs, digg relays 2026-10-18 (publication dates not individually pinned)each source page
Context events (outside window, labelled)2026-08-18 Free Mode post (GPT-5.6 Luna); 2026-08-19 OpenAI partnership page; 2026-09-29 "Free the models"; 2026-10-02 changelog (GPT-6.1 Sol + Claude Sonnet 5.5 added); 2026-10-16 changelog (Claude Haiku 5.5 added to Power Mode — inside window, day after the remarks)Replit blog/changelog, OpenAI site

--

✓

What happened?

🎓 For Explorer

FACT (CONFIRMED): On Thursday 2026-10-15, on stage at TechCrunch Disrupt 2026 in San Francisco, Replit CEO Amjad Masad was asked whether Replit's August partnership with OpenAI meant "no more Anthropic," and answered: "Yeah. I mean, yeah." TechCrunch published the exchange on 2026-10-16 at 2:22 pm PT under the headline "Replit is no longer working with Anthropic, CEO says" (https://techcrunch.com/2026/10/17/replit-is-no-longer-working-with-anthropic-ceo-says/, original /2026/10/16/… URL still indexed).

FACT (CONFIRMED): TechCrunch's article carries two further on-stage quotes: Masad saying "I think ChatGPT is better now", and — pressed on the basis for the change — "consistently, Sol and Astra perform better on building vibe-coded applications from scratch … decision based on both cost, partnership, and performance." He also described Replit as "until recently, a top-five customer of Anthropic's."

FACT (CONFIRMED — with contradiction inside the same article): TechCrunch's own reporting undercuts the headline: one person familiar with the partnership told TechCrunch the Replit–Anthropic partnership is still in place, and the article notes Replit's late-September 2026 blog post citing Claude Fable 5/5.1 running in production, plus an employee's X post about Anthropic's Opus 5.5. TechCrunch updated and effectively re-titled the piece on 2026-10-17 at 11:48 AM PDT (URL path moved to /2026/10/17/…).

FACT (COMPANY CLAIM, relayed): Also on 2026-10-17, Masad posted a written clarification: "Replit remains model agnostic … OpenAI is now our preferred partner for powering Replit Agent, based on performance, economics, and the strength of our partnership." In replies he said "OpenAI models are generally more affordable" and pointed to Replit's 2026-10-02 changelog entry that had added GPT-6.1 Sol alongside Claude Sonnet 5.5. The direct X status was not retrieved; the full text comes from two relays (zeitgraphs, WebProNews).

FACT (CONFIRMED — independent evidence): As of 2026-10-18, Replit's public documentation still offers Anthropic models in Replit Agent — the model selector lists Claude Fable 5.1, Opus 5.5, Opus 5, Opus 5 Fast and Opus 4.8 (Max plan), Sonnet 5 and Sonnet 4.6 (Power plan) — and the Replit changelog entry dated 2026-10-16 (the day after the remarks) adds Claude Haiku 5.5 to Power Mode. Anthropic's Replit customer case study remains live.

INTERPRETATION: What actually happened is narrower than the headline: Replit's CEO declared OpenAI the preferred/default partner for Replit Agent, not the termination of an engineering relationship that its own documentation still shows running. The story is as much about the correction (headline → clarification → insider pushback) as about the switch.

Δ

What changed?

Before: Replit Agent, first launched September 2024, was built on Anthropic's Claude (Claude 3.5 Sonnet at launch), and the Replit–Anthropic pairing was one of the visible poster children of the Claude-for-coding era: Anthropic still hosts a Replit case study claiming revenue grew from $1M to $240M ARR with Agent 4 running on Claude Sonnet 4.6/Opus 4.7, and Inc. documented the "after partnering with Anthropic, Replit has grown revenue by 10x" narrative in February 2025. Through mid-2026 Replit's own benchmarking (ViBench, May 2026 run) had Anthropic's Opus 4.8 marginally ahead of OpenAI's GPT-5.5 on its zero-to-one task (87.8% vs 86.5%, per Georgian's write-up).

Change (2026-08 → 2026-10-15): Replit introduced Free Mode "powered by OpenAI's GPT-5.6 Luna" (2026-08-18), then announced the broader OpenAI partnership (OpenAI's own page, 2026-08-19, with Masad and Sam Altman quotes, routing Replit Agent to GPT-5.6 Sol). At Disrupt on 2026-10-15 Masad confirmed the direction of travel in the starkest terms ("Yeah. I mean, yeah"), named the current models (Sol, Astra) as consistently better for building from scratch, and gave the tripartite basis: cost, partnership, performance. On 2026-10-17 he formalised it: OpenAI is the preferred partner for powering Replit Agent, while Replit "remains model agnostic."

After: OpenAI models are positioned as the default path for Replit Agent (Free Mode routes automatically; Sol/Astra headline the paid tiers), but Anthropic models remain selectable in the product and were being added days around the remarks. The partnership wording reversed within 24 hours ("no longer working with" → "preferred partner … model agnostic"), with an insider directly disputing the original claim. Discovery's event_date (2026-10-16) captured the headline, not the event (2026-10-15) — and the headline itself was amended by 2026-10-17.

↔

Before → Change → After

🎓 For Explorer
  • Before (2024-09 → 2026-08): Replit Agent = Claude shop. Launched on Claude 3.5 Sonnet (2024-09); the partnership powered a 10x revenue run (Inc., 2025-02-18); Anthropic's case study claims $1M → $240M ARR; ViBench's May 2026 zero-to-one run put Opus 4.8 above GPT-5.5. Anthropic was, on Masad's own Disrupt account, a recent "top-five customer."
  • Change (2026-08-18 → 2026-10-17): Free Mode launches on OpenAI GPT-5.6 Luna (Replit blog, 2026-08-18 — the discovery-listed primary); OpenAI publishes the partnership (2026-08-19) routing Replit Agent to GPT-5.6 Sol; a 2026-10-02 changelog adds GPT-6.1 Sol and Claude Sonnet 5.5 (multi-vendor still); 2026-10-15 Masad says onstage Replit is "no longer working with Anthropic" and that Sol/Astra win "based on both cost, partnership, and performance"; 2026-10-16 TechCrunch publishes the "no longer working" headline while quoting an insider who says the partnership stands; 2026-10-17 Masad clarifies in writing: preferred partner, model agnostic.
  • After (as verified 2026-10-18): Replit Agent's public model selector still lists Claude Fable 5.1/Opus 5.5/Opus 5/Opus 5 Fast/Opus 4.8 (Max), Sonnet 5/Sonnet 4.6 (Power) beside GPT-6 Sol, GPT-6 Luna Fast (Power), GPT-6 Astra (Max) and Kimi K3 (Max); the 2026-10-16 changelog adds Claude Haiku 5.5; no official Replit blog post announces any switch (blog index checked 2026-10-18); Anthropic's case study remains live. Default routing toward OpenAI = COMPANY CLAIM; continued Anthropic availability = INDEPENDENTLY VERIFIED.
⚙

How it works

5.1 What "default model" means in Replit's architecture

Replit Agent is a hosted agentic coding environment with tiered plans (Free / Power / Max, plus Auto, and a single Auto mode for Enterprise per the agent-modes docs). Two mechanisms matter:

  1. Routing (Free Mode and Auto): Free Mode has no manual model picker — the platform routes requests. Replit's Free Mode post says it is "powered by OpenAI's GPT-5.6 Luna" and states "With OpenAI, Replit now helps you accomplish even more." In routing-tier products, "default model" is literally an infrastructure configuration: this is where Masad's claim is most plausibly true, and also where it is least externally verifiable — no public endpoint exposes the routing table.

  2. Selection (paid tiers): In Power and Max plans the user picks from a model selector. As verified in the lab on 2026-10-18 (https://docs.replit.com/features/agent/model-selector), that selector includes:

    • Max: Claude Fable 5.1, Claude Opus 5.5, Claude Opus 5, Claude Opus 5 Fast, Claude Opus 4.8, GPT-6 Astra, Kimi K3
    • Power: Claude Sonnet 5, Claude Sonnet 4.6, GPT-6 Sol, GPT-6 Luna Fast (plus Claude Haiku 5.5 added 2026-10-16 per changelog)

    FACT (INDEPENDENTLY VERIFIED): Anthropic models are not merely present but are being added concurrent with the "no longer working" statement.

5.2 The harness is deliberately model-agnostic

FACT (COMPANY CLAIM, Replit engineering blog): "Free the models" (2026-09-29) describes Replit's stack with GPT-6 Astra in the core Max loop while running Claude Fable 5/5.1 in production, praising both, and frames the harness as model-agnostic by design. This predates the Disrupt remarks by two weeks and is the technical reason a "switch" is a routing/partnership decision rather than a rewrite — and why "model agnostic" in the clarification is architecturally credible even if the preferred-partner designation is real.

5.3 Why Replit says it moved: cost, partnership, performance

  • Performance (COMPANY CLAIM): "consistently, Sol and Astra perform better on building vibe-coded applications from scratch"; "I think ChatGPT is better now."
  • Cost (COMPANY CLAIM): Masad's 2026-10-17 reply: "OpenAI models are generally more affordable."
  • Partnership (COMPANY CLAIM): "the strength of our partnership" (OpenAI) — formalised 2026-08-19 with a joint announcement page carrying Masad and Sam Altman quotes.
  • INDEPENDENT EVIDENCE check (§8): none of the three is corroborated by retrievable independent benchmark data for the current model generation; the last retrievable public figures (ViBench, May 2026) pointed the other way.
!

Why it matters

🎓 For Explorer

FACT/INTERPRETATION: Replit is one of the highest-visibility agentic-coding products, and its model stack has been a Claude showcase since 2024. A preferred-partner designation from Replit to OpenAI — whatever the contractual fine print — is a demand-side signal in the model-provider contest for agentic software engineering: the most-quoted "Claude company" is now publicly saying GPT-class models win on cost and performance for building from scratch, at least for its workload. It matters equally for what it shows about switching costs in agent harnesses: because Replit's harness is model-agnostic by design, the supplier change was a configuration and commercial act, not an engineering migration — which is precisely the leverage every model vendor now contests with agent-platform defaults. And it matters because the correction cycle (headline → insider contradiction → CEO clarification inside 48 hours) demonstrates that "vendor X is out" claims in this market now require artifact-level verification — as this story's lab does.

✦

What became possible?

🎓 For Explorer

PREDICTION/INTERPRETATION grounded in evidence:

  1. Multi-vendor defaulting becomes the norm, not defection: Replit's own 2026-10-02 changelog added GPT-6.1 Sol and Claude Sonnet 5.5 in the same entry; Anthropic models remain in all paid tiers. "Preferred partner" likely means routing defaults, commercial terms and co-marketing — PREDICTION: other agent platforms will announce similar "preferred partner" language while keeping every vendor selectable.
  2. Benchmark diplomacy: if Replit's performance claims are to be believed, expect Replit to publish fresh ViBench numbers for Sol/Astra vs Claude generation models — the existing public leaderboard (last retrievable figures May 2026) currently contradicts the claim.
  3. Anthropic-side response options: public case study stays up (as of 2026-10-18) — PREDICTION: Anthropic either updates it quietly or Replit archives it; a loud termination would require Replit to remove Claude from docs, which has not happened.
  4. Procurement precedent for enterprises (§10): a vendor that switches frontier defaults mid-flight without a product migration gives cover to enterprises treating coding agents as swappable model layers.
◎

Implications

Technical

  • Defaults are the real battleground. In routed tiers (Free Mode, Auto), the model choice is invisible to the user and fully controlled by the platform. INTERPRETATION: Masad's operational claim — OpenAI as default — is most meaningful exactly where it is least testable from outside; only the paid-tier selector is public evidence, and it shows inclusion, not preference.
  • Replit's public benchmark does not currently support the performance claim. Replit's ViBench ("our public benchmark for vibe coding," https://replit.com/blog/evaluating-and-improving-agent-at-scale, 2026-06-23; paper: https://dl.acm.org/doi/10.1145/3786335.3813162) had, per Georgian's 2026-06-25 write-up, a May 2026 zero-to-one comparison of Claude Opus 4.8 at 87.8% vs GPT-5.5 at 86.5% — Anthropic ahead — with Replit running ViBench roughly 10× per week. COMPANY CLAIM status attaches to "Sol and Astra perform better": no retrievable post-switch leaderboard result corroborates it, and the live leaderboard (https://vibench.ai/) failed to fetch during this run (see §13).
  • Model-agnostic harness as strategic asset: "Free the models" (2026-09-29) shows Astra in the core loop and Fable 5/5.1 in production simultaneously. The technical implication for the industry: agent harness quality, not model lock-in, is the moat — supplier switches become negotiation levers rather than migrations.
  • Version churn is the confounder: between 2026-10-02 and 2026-10-16 Replit's changelog added GPT-6.1 Sol, Claude Sonnet 5.5 and Claude Haiku 5.5. Any "X beats Y" claim at Disrupt describes moving targets; the docs snapshot of 2026-10-18 (five Claude models in Max, two in Power — three in Power after the 2026-10-16 addition) is the only externally fixed evidence.

Developer

  • Nothing breaks today. Claude models remain selectable in Replit Agent's Power/Max tiers (verified 2026-10-18). Builders on Claude via Replit need no immediate migration; check the selector, not the headlines.
  • Free Mode users have no choice by design. If you depend on a specific model behaviour on the free tier, you do not control it — routing is platform-controlled (documented in the model-selector docs). Move to a paid tier for explicit model selection.
  • Treat vendor-swap statements as configurations, not migrations: Replit's architecture shows the pattern to copy — keep the harness model-agnostic, keep every provider selectable, let commercials and routing decide defaults. If you build agent products, this is your leverage with model vendors.
  • Verify with artifacts: for this class of story, the model-selector docs, the changelog and the vendor's own announcement page are the primary evidence; news headlines (including the re-titled TechCrunch piece) lag and simplify.
  • Cost claims deserve your own measurement: Masad's "generally more affordable" (COMPANY CLAIM) is workload-specific; benchmark on your own agent traces before re-architecting.

Enterprise

  • For Replit customers: enterprise tier runs a single Auto mode (per agent-modes docs) — meaning Replit controls routing entirely for enterprise deployments. Enterprises that assumed a Claude-standard stack should now ask Replit contractually which models power Auto, since the docs do not disclose it and the CEO's statements are the only public signal. The CEO also disclosed (per TechCrunch) enterprise rising from 2% → 15% of revenue, B2B >50% of revenue, and >50% enterprise + government expected by 2027 — figures that make default-model messaging a procurement question, not a developer preference. (All COMPANY CLAIM.)
  • For Anthropic: the visible "Claude built Replit" narrative ($1M → $240M ARR case study, still live) is now awkward; expect Anthropic to lean on other coding partnerships.
  • For OpenAI: the August partnership page (Masad + Altman quotes) plus "preferred partner" converts a competitor's showcase into a distribution win — with the caveat that "preferred" coexists with Claude, GPT and Kimi in the selector.
  • For buyers of coding agents generally: INTERPRETATION: model concentration risk in coding platforms is now demonstrable — one CEO statement can re-point a platform's default. Contractual model-portability (the ability to pin or whitelist models) is the mitigating control to demand.

Strategic

  • The "Claude company" era is ending as a category claim. Anthropic's coding moat was partly built on Replit-shaped showcases; if Replit's preferred partner is OpenAI, the flagship reference rotates. INTERPRETATION: OpenAI's August push (partnership page, GPT Sol/Astra branding aimed at agent workloads) is landing on real defaults.
  • Cost may be the decisive axis (COMPANY CLAIM): Masad's two-part justification — better and cheaper — implies frontier coding workloads are becoming price-competitive at the top, which pressures every vendor's inference margins and favours platforms that can route arbitrage.
  • "Model agnostic" is the new strategic line: Replit's clarification shows platforms asserting neutrality while taking partner money. PREDICTION: preferred-partner arrangements become the standard commercial shape — platform keeps optionality, vendor gets default placement.
  • Replit's IPO posture: the Unite AI relay (https://www.unite.ai/; caveat: AI-generated byline) reports Masad discussing an IPO intent and NRR >400% at Disrupt; if so, the OpenAI alignment reads as a marquee partnership ahead of a listing (EARLY RESEARCH — single low-independence source).
  • Equity, not contracts, was checked: no OpenAI or Anthropic equity in Replit was found; Replit's investors include a16z, Coatue, Prysm, Georgian, Google AI Futures Fund, QIA and Visa (per revenuememo's 2026-09-30 funding profile; Dealroom profile consistent), with Greg Brockman as a 2018 individual angel predating this switch. INTERPRETATION: the switch is commercial (cost/performance/partnership as Masad states), not equity-driven.
⚠

Risks & limitations

Risks
  • The claim may be overstated and is already hedged: the original headline ("no longer working with") was contradicted within publication by an insider and walked back within ~34 hours ("preferred partner… model agnostic"). Anyone planning migrations off the headline would have acted on a false premise.
  • Anthropic relationship risk for Replit: publicly demoting a supplier that still powers documented production paths invites negotiation friction; the still-live Anthropic case study is a hostage of sorts.
  • Single-insider dependency: the "partnership still in place" pushback is one unnamed person; if wrong, the continued Claude availability is merely short-term product inertia. Both readings remain live — the core dispute is unresolved (digg's relay makes the same point: neither Replit nor Anthropic has publicly resolved the contradiction).
  • Benchmark credibility risk for Replit: making performance claims its own public benchmark last contradicted invites scrutiny of forthcoming ViBench runs; selective publishing would damage Replit's evaluation credibility.
  • Concentration risk: defaulting to one vendor's models in routed tiers (Free Mode/Auto) ties product quality to one supplier's rate limits, deprecations and pricing moves.
  • Narrative whiplash for the ecosystem: Claude-coding tutorials, courses and enterprise assumptions built on Replit now carry an expiry question — an education-market risk as much as an engineering one.
Limitations
  • Direct X post not retrieved: Masad's 2026-10-17 clarification text is known only through two relays (zeitgraphs, WebProNews) that agree with each other; the direct X status URL was not fetched successfully. Treated as PRESS REPORT (relay) of a company statement.
  • Session video not transcribed: the Disrupt session recording (https://www.youtube.com/watch?v=grkRhVKcxiE) renders via JavaScript and was not transcribed directly; all on-stage quotes come from TechCrunch's report (the outlet that ran the session) plus Unite AI's relay. High confidence, but relay-level.
  • Unite AI (https://www.unite.ai/) uses an AI-generated byline — used only for quote corroboration and clearly flagged context (VBench internal-benchmark reference, NRR >400%, IPO remarks); never sole support for a load-bearing claim.
  • The Vibench leaderboard (https://vibench.ai/) failed to fetch (transport error/empty response) during this run — the "current standings" question therefore rests on the June 2026 Georgian/Replit write-up of May 2026 figures, not a live reading. The ACM paper page was located but not used for scores.
  • No official OpenAI or Anthropic statement exists in anything retrieved — neither vendor commented on the switch in any fetched source. Absence noted, not inferred.
  • No official Replit announcement of the switch: the Replit blog index (checked 2026-10-18) has no post about the default change; the story exists only as CEO remarks + his X clarification. This is why the substance stays COMPANY CLAIM.
  • Routing internals unverifiable: Free Mode/Auto routing is server-side; the lab cannot observe which model answers a given request.
  • (a) event_date 2026-10-16 is publication, not the 2026-10-15 event; (b) the listed primary source (Free Mode post) never mentions Anthropic or a default switch (§14 Q1). Both flagged here per evidence discipline; was not modified.
  • Publication dates not individually pinned for Unite AI, zeitgraphs and digg ( 2026-10-18); their content is used for corroboration, not chronology.
  • Low-quality AI aggregators restating the story surfaced in search and were deliberately excluded (same policy as other stories in this run).

--

?

Open questions

  1. Why did discovery's primary source not support the claim? The Free Mode post (2026-08-18) evidences only the August OpenAI partnership; the default-switch claim traces entirely to the 2026-10-15 Disrupt remarks. Was a Replit changelog or docs diff intended as primary evidence?
  2. What does "preferred partner" mean contractually — revenue share, exclusive routing in Free/Auto, co-marketing, or benchmark collaboration? Nothing public specifies.
  3. Does the Replit–Anthropic partnership still exist in writing, as the unnamed insider says — and if so, on what terms and for which tiers?
  4. When Replit's changelog added GPT-6.1 Sol (2026-10-02) and Claude Haiku 5.5 (2026-10-16) — which models does Auto/Enterprise routing actually pick, and can enterprise customers pin them?
  5. Will Replit publish fresh ViBench results for Sol/Astra vs the current Claude generation? The last public comparison favoured Anthropic.
  6. Did Anthropic respond at all? No statement found in any retrieved source.
  7. Does Masad's "top-five customer of Anthropic's" hold for 2026, and did Anthropic know before Disrupt?
  8. What is the fate of the Anthropic case study (live as of 2026-10-18) — updated, archived, or left to age?
↗

What happens next?

🎓 For Explorer
  • Near-term (days–weeks): Watch for (a) any Replit blog/changelog statement actually announcing the default change (none as of 2026-10-18); (b) an Anthropic statement (none found); (c) Masad's X thread gaining an official correction or confirmation; (d) TechCrunch further amending the piece (already once updated, 2026-10-17 11:48 AM PDT).
  • Medium-term (weeks–quarters): PREDICTION: a fresh Replit ViBench publication positioning Sol/Astra against the current Claude generation (to substantiate the Disrupt claim); Anthropic either refreshing its Replit case study or it quietly disappearing; "preferred partner" framing spreading to other agent platforms as the commercial template; Replit's next changelog entries revealing the next multi-vendor additions (pattern: 2026-10-02 added both GPT-6.1 Sol and Claude Sonnet 5.5).
  • Watch metric: whether Claude disappears from any Replit paid tier's selector. Its removal — not a CEO soundbite — would be the first independently verifiable evidence that the relationship actually ended. As of 2026-10-18: present in both Power and Max, with Haiku 5.5 newly added.
★

Editorial takeaway

🎓 For Explorer

Replit's CEO stood on the TechCrunch Disrupt stage on 2026-10-15 and said what the aggregated headline ran with: "Yeah. I mean, yeah" — no more Anthropic. Within a day TechCrunch's own source contradicted it, and within 34 hours Masad had rewritten it as "OpenAI is now our preferred partner… Replit remains model agnostic." The artifacts side with the clarification, not the headline: five Claude models sit in Replit's Max tier, two more in Power, and the day after the remarks Replit's changelog added Claude Haiku 5.5. What is confirmed is the preference; what is disproven is the divorce. The real story is that the most famous "Claude company" now says GPT wins on cost and performance for building from scratch — a COMPANY CLAIM resting on a benchmark whose last published standings favored Anthropic — while its model-agnostic harness lets it say so without breaking anything. In agentic coding, the supplier war is fought in routing tables, and only one side's press release is visible.

Lab: VERIFY

Objective:

A claim of supplier disintegration is artifact-falsifiable: if the "former" supplier's models are still documented, selectable, and being added to the product days around the statement, the literal claim fails while the weaker "preferred partner" claim may still stand. The lab therefore asks, for each artifact:

  • Does Replit's own product surface still carry Anthropic models, and did it change around 2026-10-15?
  • Does any Replit first-party page actually announce an OpenAI default?
  • Does the discovery-listed primary source support the story as framed?
STEP 1 — Check the live model selector (the decisive artifact)

Fetched https://docs.replit.com/features/agent/model-selector (2026-10-18).

Result:

TierModels listed (observed)
MaxClaude Fable 5.1, Claude Opus 5.5, Claude Opus 5, Claude Opus 5 Fast, Claude Opus 4.8, GPT-6 Astra, Kimi K3
PowerClaude Sonnet 5, Claude Sonnet 4.6, GPT-6 Sol, GPT-6 Luna Fast (+ Claude Haiku 5.5 per STEP 2)
Free ModeNo manual model choice — platform routing only (documented as automatic)

Finding: Seven Claude model entries across the paid tiers (five in Max, two in Power) — eight counting the Haiku 5.5 addition in STEP 2 — and none removed. "No longer working with Anthropic" fails this check as a literal product claim. "OpenAI is now our default" is not directly checkable here (selection ≠ routing), but Free Mode's documented automatic routing is the mechanism where a default would live — unobservable from outside.

STEP 2 — Check the changelog around the remarks (the timing test)

Fetched https://replit.com/changelog (2026-10-18).

Result:

Changelog entryDateContent relevant to the claim
GPT-6.1 Sol and Claude Sonnet 5.52026-10-02Both vendors added in the same entry — two weeks before "no longer working with Anthropic"
Claude Haiku 5.5 added to Power Mode2026-10-16The day after the remarks, Replit added an Anthropic model

Finding: The product timeline around the statement moves toward Anthropic, not away. A company genuinely ending a relationship does not ship the partner's new model the next day. Strong artifact-level contradiction of the headline claim; consistent with the insider's "partnership still in place."

STEP 3 — Check for an official Replit announcement of the switch

Fetched https://replit.com/blog (index, 2026-10-18) and scanned for any default-switch post.

Result: No official Replit post announcing a default switch to OpenAI exists on the blog index. The entire story rests on (a) on-stage remarks at TechCrunch Disrupt (2026-10-15) and (b) Masad's X clarification of 2026-10-17 (relayed; direct status not retrieved).

Finding: The substance — "OpenAI is our default/preferred partner" — has no first-party written artifact at time of run. COMPANY CLAIM ceiling confirmed; discovery's CONFIRMED status cannot be inherited by the substance.

STEP 4 — Check whether the relationship is still publicly presented by Anthropic

Fetched https://www.anthropic.com/customers/replit (2026-10-18).

Result: Case study live and unchanged in substance (Agent 4 on Claude Sonnet 4.6 / Opus 4.7; ARR claim $1M → $240M). No takedown, no caveat banner.

Finding: The counterparty still presents Replit as a customer case. Neither side has published a termination — matching digg's observation that neither Replit nor Anthropic has publicly resolved the contradiction. (Absence of a takedown is evidence of continuity, not proof of contract terms.)

STEP 5 — Verify the discovery-listed primary source actually supports the story

Fetched https://replit.com/blog/replit-introduces-free-mode (2026-08-18).

Result: The post is about Free Mode, "powered by OpenAI's GPT-5.6 Luna," with "With OpenAI, Replit now helps you accomplish even more." It never mentions Anthropic, a default model, or any switch; it predates the Disrupt remarks by two months.

Finding: Discovery's primary source supports only the August OpenAI partnership (corroborated independently by https://openai.com/index/replit/, 2026-08-19 — fetched, partnership confirmed with Masad + Altman quotes). The default-switch claim traces exclusively to the 2026-10-15 remarks. Recorded as a discovery-evidence mismatch §13/§14.

STEP 6 — Attempt the public benchmark leaderboard (performance-claim check)
ProbeResult
https://vibench.ai (leaderboard)Fetch failed — transport error / empty response, 2026-10-18
https://replit.com/blog/evaluating-and-improving-agent-at-scale (ViBench origin, 2026-06-23)Fetched — ViBench described as "our public benchmark for vibe coding"
https://georgian.io/posts/georgian-replit-vibench (2026-06-25)Fetched — last retrievable comparison: May 2026 zero-to-one Opus 4.8 87.8% vs GPT-5.5 86.5% (Anthropic ahead then); ViBench runs ~10×/week

Finding: Masad's "consistently, Sol and Astra perform better" is uncorroborated by any retrievable public data — the last public standings point the other way and are now four+ months stale. Claim remains COMPANY CLAIM; the lab could neither confirm nor fairly falsify it (stale data, live leaderboard unreachable).

STEP 7 — Confirm the event date against discovery

Checked https://techcrunch.com/events/techcrunch-disrupt/agenda/ (event 2026-10-13 → 2026-10-15), the 2026-10-16 session recap ("on Thursday"), and the WebProNews relay ("Thursday, Oct. 15").

Result: On-stage event = 2026-10-15; discovery's event_date = 2026-10-16 (the article's publication date). Both inside the configured window (2026-10-14 → 2026-10-17), so window validation passes either way; discrepancy recorded in §0.

Verdict

What the lab verified: (1) Anthropic models are still shipped, selectable, and newly added (Claude Haiku 5.5 on 2026-10-16) in Replit Agent as of 2026-10-18 — the literal "no longer working with Anthropic" claim fails against live artifacts, while the insider report and Masad's own "model agnostic" clarification stand corroborated. (2) No first-party Replit announcement of any switch exists — the substance is capped at COMPANY CLAIM. (3) Anthropic still publicly presents Replit as a case study. (4) Discovery's listed primary supports only the August partnership, not the switch. (5) The performance claim has no retrievable independent support; last public figures favoured Anthropic (May 2026).

What the lab could not test: actual routing defaults in Free Mode/Auto/Enterprise tiers (server-side, unobservable); whether a preferred-partner contract exists (no public document); Masad's X post directly (relays used); current ViBench standings (leaderboard unreachable); the session video transcript (JS-only page; TechCrunch's own reporting used).

One-line takeaway: Replit's docs are open the day after its CEO said the Anthropic relationship was over — five Claude models in Max, two in Power, one brand-new in the changelog — so what actually changed is who is preferred, not who is gone.

Research sources

Primary Sources (14)
Primary
Anthropic — Replit customer case study**the Anthropic–Replit relationship remains publicly presented by Anthropic** — Agent 4 on Claude Sonnet 4.6/Opus 4.7, ARR claim $1M→$240M (COMPANY CLAIM). Still live days after "no longer working with Anthropic." — counterparty primary; live-artifact evidence. ---
Visit source ↗
Primary
ACM DL — ViBench paper (DOI page)existence of the published ViBench benchmark paper (research standing of Replit's evaluation). Scores not extracted from this page. — primary research record.
Visit source ↗
Primary
Replit — "Free the models"harness is model-agnostic by design; GPT-6 Astra in the core Max loop **and** Claude Fable 5/5.1 in production simultaneously (COMPANY CLAIM of engineering context, two weeks pre-Disrupt). — company technical blog.Date: 2026-09-29
Visit source ↗
Primary
Replit — "Evaluating and improving Replit Agent at scale" (ViBench)ViBench as "our public benchmark for vibe coding"; Replit's benchmarking practice (context for §8). — company technical blog.Date: 2026-06-23
Visit source ↗
Primary
Replit — Blog index**no official Replit post announcing any default switch** exists as of retrieval — the story lives only in CEO remarks + X clarification. — negative primary check.
Visit source ↗
Primary
Replit — Changelogentry dated **2026-10-16 (day after the remarks) adds Claude Haiku 5.5** to Power Mode; 2026-10-02 entry added GPT-6.1 Sol **and** Claude Sonnet 5.5 (per Masad's relayed reply) — concurrent multi-vendor additions around the "switch." — company primary, dated, artifact-level.
Visit source ↗
Primary
Replit Docs — Agent modesFree / Power / Max plans plus Auto; Enterprise runs a single Auto mode (routing control point for enterprise, research §10). — live product documentation.
Visit source ↗
Primary
Replit Docs — Agent model selector**INDEPENDENTLY VERIFIED continued Anthropic integration** — Max tier lists Claude Fable 5.1, Opus 5.5, Opus 5, Opus 5 Fast, Opus 4.8 alongside GPT-6 Astra and Kimi K3; Power tier lists Claude Sonnet 5, Sonnet 4.6 alongside GPT-6 Sol, GPT-6 Luna Fast; Free Mode routes automatically with no manual model choice. — live product documentation — the strongest artifact-level evidence in this story.
Visit source ↗
Primary
OpenAI — "Replit" partnership announcementthe August partnership is real (INDEPENDENT EVIDENCE of the relationship): Masad and Sam Altman quotes, Replit Agent routing to GPT-5.6 Sol; site nav shows GPT-6.1 Sol / GPT-6 Astra branding. — counterparty primary source.Date: 2026-08-19
Visit source ↗
Primary
Replit — "Replit introduces Free Mode" (discovery-listed primary)**only the August OpenAI partnership** — Free Mode "powered by OpenAI's GPT-5.6 Luna," "With OpenAI, Replit now helps you accomplish even more." **Never mentions Anthropic or a default switch** — recorded as a discovery-evidence mismatch (research §13/§14 Q1). — company primary; scope limited to what it actually says.Date: 2026-08-18
Visit source ↗
Primary
TechCrunch Disrupt session recording — "the self-driving company" (YouTube)- Fetch outcome: **page renders JavaScript-only; transcript not extracted in this run.** Recorded as the existence of primary video for the remarks. — primary pointer; not transcribed — quotes used here are TechCrunch's (outlet that ran the session).
Visit source ↗
Primary
TechCrunch Disrupt 2026 — official event agenda pageevent window and venue; bounds the possible event dates to 2026-10-13/14/15, corroborating the "Thursday 2026-10-15" dating. — primary venue/chronology.Date: event ran 2026-10-13 → 2026-10-15, Moscone West, San Francisco
Visit source ↗
Primary
TechCrunch — "Replit's CEO says the future of work is the self-driving company" (session recap)fixes the actual event date (2026-10-15, i.e. the final day of Disrupt) against discovery's 2026-10-16; session context. — primary for event chronology.Date: 2026-10-16 (recap states remarks were made "on Thursday" = 2026-10-15)
Visit source ↗
Primary
TechCrunch — "Replit is no longer working with Anthropic, CEO says" *(the story's origin article; byline not relied upon in this research)*the on-stage exchange of 2026-10-15 — Masad's "Yeah. I mean, yeah"; "I think ChatGPT is better now"; "consistently, Sol and Astra perform better on building vibe-coded applications from scratch … decision based on both cost, partnership, and performance"; "until recently, a top-five customer of Anthropic's"; the written clarification "Replit remains model agnostic… OpenAI is now our preferred partner for powering Replit Agent, based on performance, economics, and the strength of our partnership"; **the insider contradiction** (one person familiar says the partnership is still in place); the late-September Replit blog citing Claude Fable 5/5.1 in production and an employee post on Opus 5.5; enterprise metrics (2%→15% of revenue, >50% B2B, >50% enterprise+gov expected 2027). — origin primary reporting; quotes are relay-level for the session itself (the article IS the first report of the on-stage remarks).
Visit source ↗
Independent Sources (7)
Independent
Dealroom — Replit (repl-it) company profilecorroboration of the cap-table/investor picture (no OpenAI or Anthropic stake surfaced). — independent data provider. ---
Visit source ↗
Independent
Revenue Memo — Replit funding profileinvestor list (a16z, Coatue, Prysm, Georgian, Google AI Futures Fund, QIA, Visa — May 2026 round) and Greg Brockman's 2018 individual angel investment — used to assess equity-vs-commercial framing of the switch (no OpenAI/Anthropic corporate equity found). — independent funding reporting.
Visit source ↗
Independent
Georgian — "Georgian + Replit: ViBench"**the only retrievable quantitative comparison** — May 2026 zero-to-one: Claude Opus 4.8 **87.8%** vs GPT-5.5 **86.5%** (Anthropic ahead then); Replit runs ViBench ~10×/week. Basis for rating the Disrupt performance claim COMPANY CLAIM, not verified. — independent technical write-up of Replit's benchmark (investor firm with Replit relationship — noted).
Visit source ↗
Independent
Inc. (Ben Sherry) — "After Partnering With Anthropic, Replit Has Grown Revenue by 10X"prior Anthropic-centric relationship — Replit Agent launched September 2024 on Claude 3.5 Sonnet; 10x revenue narrative; scale of what "switching" reverses. — independent business reporting, historical context.
Visit source ↗
Independent
WebProNews — "Replit names OpenAI its preferred partner, making ChatGPT the default over Anthropic's Claude"second independent relay of the 2026-10-17 clarification wording; independent confirmation of the **2026-10-15 event date**. — independent relay.
Visit source ↗
Independent
zeitgraphs — "Replit names OpenAI preferred partner while staying model agnostic"full text of Masad's written clarification ("OpenAI is now our preferred partner… model agnostic"), plus his reply "OpenAI models are generally more affordable" and the reference to the 2026-10-02 changelog adding GPT-6.1 Sol + Claude Sonnet 5.5. — independent relay of the company clarification (direct X status not retrieved).
Visit source ↗
Independent
Unite AI () — "Replit CEO Masad Says Company No Longer Works With Anthropic"verbatim relay of the on-stage Q&A — the interviewer's "no more Anthropic?" question and Masad's **"Yeah. I mean, yeah"**; links to the session recording; context (internal benchmark reference, NRR >400%, IPO remarks — flagged EARLY RESEARCH). — **corroborating relay with caveat: AI-generated byline** — never sole support for a load-bearing claim.
Visit source ↗
Secondary Sources (1)
Secondary
digg — Replit/Anthropic contradiction coverageexplicit statement that **"Neither Replit nor Anthropic has publicly resolved the contradiction"**; relays Yahoo Finance and WebProNews coverage of the same exchange (plus links to two further tech-news relays, not fetched by this research). — secondary aggregator of the coverage cluster; used only for the meta-point that the contradiction stands unresolved publicly. ---
Visit source ↗
Up next · story 6 of 10
TSMC posts sixth straight record quarterly profit on AI chip demand
All storiesNext story