News Weekly
ENहिन्दी
LV 10 XP
0% read
Your progress · 0/5 chapters
About 15 min total
Models & EconomicsISSUE #6 · STORY 3 OF 20Oct 7, 2026CONFIRMED

Anthropic cuts Haiku pricing 90% and puts a 1M-token context on its cheapest tier

On 7 October 2026 Anthropic released Claude Haiku 5.5 at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens, a 90% per-token cut versus Haiku 4.5 that matches OpenAI's GPT-6 Luna exactly. The per-task saving is nearer 75% because the new tokenizer emits about 30% more tokens for the same text.

A horizon-long line of identical small workstations, each with the same cheap worker, beneath two rival price banners falling past each other.

Read it your way

CHAPTER 1Picked for Explorer

The 60-second version

On 7 October 2026 Anthropic shipped Claude Haiku 5.5, the third model in the Claude 5.5 family, at $0.10 input and $0.50 output per million tokens for prompts up to 100,000 tokens. That matches OpenAI's GPT-6 Luna exactly and reverses Haiku 4.5's former position as the most expensive small model.

A 90% per-token cutThe sub-100K band fell from $1.00 and $5.00 on Haiku 4.5 to $0.10 and $0.50.
A price cliff above 100K tokensPrompts over 100,000 tokens jump to $0.50 and $2.50, so long-context work costs far more than the headline.
Context rises to 1M tokensThe window grew from 200,000 tokens, paired uniquely with two-tier prompt-length pricing.
An effort dial comes to HaikuHaiku 5.5 accepts a Low-to-Max effort control, the first time a Haiku-class model has one.
Finish this chapter for +15 XP
A cargo run crossing a marked band, then climbing a hard step to a dearer band, while the harbour's measuring marks have been redrawn closer.
Your next move · Explorer

Re-run your eval before trusting the cut

1Migrate your highest-volume Claude workload to the Haiku 5.5 model ID and re-run your own eval suite.
2Measure cost per task and latency, and watch whether any prompt crosses the 100,000-token band.
3Set the effort level per task class rather than leaving the default, which scores well below max.

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 Workhorse Pricing badge

Did it stick?

0/3
What is the per-token price cut for prompts up to 100,000 tokens?+20 XP
Why is Anthropic's '75% lower average' different from the 90% per-token cut?+20 XP
Which claim is a COMPANY CLAIM rather than independently verified?+20 XP
Your call · +5 XP

Haiku 5.5 now matches GPT-6 Luna on price. What should most teams do first?

Research dossierThe full research, labeled and sourcedCONFIRMED17 sources · 15 min read · 20 sections
Story identity
FieldValue
Story IDS03
TitleAnthropic ships Claude Haiku 5.5 at roughly 75% lower cost than Haiku 4.5
OrganizationAnthropic
CategoryModel Release / Pricing
Event date2026-10-07 (announcement date 2026-10-07; article dates 2026-10-07/08)

Evidence labels used below: FACT (independently verifiable), COMPANY CLAIM (Anthropic-stated, not independently reproduced), INDEPENDENT EVIDENCE (third-party verification), INTERPRETATION (analyst reading), PREDICTION (forward-looking).

✓

What happened?

🎓 For Explorer

FACT: On October 7, 2026, Anthropic released Claude Haiku 5.5 (model ID claude-haiku-5-5), the third model in the Claude 5.5 family after Opus 5.5 (September 22) and Sonnet 5.5 (September 28). It is available on the Claude API and all major cloud platforms including Amazon Web Services (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry).

FACT: The headline pricing is $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens — a 90% per-token price cut versus Haiku 4.5 ($1.00 / $5.00). For prompts over 100,000 tokens the rates rise to $0.50 / $2.50 (50% below Haiku 4.5). Cache reads are $0.01 / $0.05 and cache writes $0.125 / $0.625 per million tokens per tier. The Batch API halves all rates ($0.05 / $0.25 within the 100K band).

COMPANY CLAIM: Anthropic states that "on average, it now costs around 75% less to run" than Haiku 4.5. The footnote explains the arithmetic: 90% lower price for requests up to 100K tokens (about 90% of Haiku 4.5 requests), 50% lower above 100K, adjusted for a new tokenizer that uses slightly more tokens per task. Independent reporting (Reuters, Yahoo Finance) repeats the "75% less" average framing.

FACT: The release was accompanied by two value moves: cache-read pricing for Claude Sonnet 5.5 was halved from $0.20 to $0.10 per million tokens (Anthropic claims ~20% cheaper on most agentic work), and a monthly Claude Platform API credit was introduced for Max and Team subscribers ($100 for Max 5x, $200 for Max 20x, up to $500 pooled for Team).

FACT: Anthropic updated its Python and TypeScript SDKs with beta support for computer use and browser use, explicitly marketed against Haiku 5.5's speed/price combination.

INDEPENDENT EVIDENCE: Reuters (2026-10-07) confirmed the launch, the pricing tiers, the "75% less" framing, the third-in-family status, and that it is Anthropic's "first Haiku with built-in safeguards for a narrow set of high-risk cybersecurity requests." Artificial Analysis independently evaluated the model (Intelligence Index 43 at max effort vs 56 for Sonnet 5.5; ~240 output tokens/sec at max, fastest in the family; AA-Briefcase v1.1 Elo 1578). Vals AI independently benchmarked it (Vals Index 54.31%, rank 16/45; Terminal-Bench 4.0 35.35% in its harness; Vibe Code Bench v1.1 90.44%, rank 3/110).

Δ

What changed?

The change is two-sided, which is why the story is significant beyond a routine refresh:

  1. Price collapse on the small tier. The per-token price fell 90% for the dominant request class (<100K prompts), reversing Haiku 4.5's position as the most expensive small model relative to OpenAI's GPT-6 Luna ($0.10/$0.50, released September 22). Simon Willison had flagged in September that Haiku 4.5 at $1/$5 was uncompetitive against Luna's $0.10/$0.50; Haiku 5.5 closes that gap at parity and, per Anthropic's benchmark table and independent evals, above-parity quality on several tests.
  2. Capability jump on the small tier. COMPANY CLAIM: OSWorld 2.1 (offline subset) 72.4% vs 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna; Terminal-Bench 4.0 39.2% vs 0.0% / 16.4%; Humanity's Last Exam 45.9% (no tools) and 57.4% (with tools) vs 10.2% / 18.7%; GDPval-AA v2.1 Elo 1620 vs 735; AA-Briefcase v1.1 Elo 1578 vs 614. These are Anthropic-reported figures; independent measurements (Artificial Analysis Terminal-Bench 33%, Vals AI 35.35% in its own harness) confirm the direction and rough magnitude.

RELATED CHANGES (FACT): adjustable effort setting arrives on a Haiku-class model for the first time (Low→Max, trading cost against intelligence per request); context window rises to 1M tokens (from 200K on Haiku 4.5) but with a two-tier prompt-length pricing structure unique among current Claude models; new tokenizer (shared with Sonnet 5.5 / Opus 5.5) produces roughly 30% more tokens for the same text, which affects per-task cost comparisons.

A cargo run crossing a marked band, then climbing a hard step to a dearer band, while the harbour's measuring marks have been redrawn closer.
↔

Before → Change → After

🎓 For Explorer
DimensionBefore (Haiku 4.5, Oct 2025 – Oct 2026)Change (2026-10-07)After (Haiku 5.5)
Price (<100K prompts)$1.00 in / $5.00 out per MTok−90% per-token price cut$0.10 in / $0.50 out per MTok
Price (>100K prompts)$1.00 / $5.00 (flat)−50%$0.50 / $2.50
Cache read$0.10 / MTok−90%$0.01 / MTok (<100K band)
Average cost to run—"~75% less on average" (COMPANY CLAIM, tokenizer-adjusted)~75% lower
Context window200K tokens5x1M tokens (two-tier pricing)
Effort controlNone on HaikuNewAdjustable Low→Max
Computer use (OSWorld 2.1 offline, claim)15.7%+56.7 pts72.4% (vs 48.9% Luna)
Terminal agentic coding (Terminal-Bench 4.0, claim)0.0%+39.2 pts39.2% (vs 16.4% Luna)
TokenizerOlder (fewer tokens per task)New~30% more tokens for same text
Cyber safeguardsHaiku 4.5 baselinesStrengthenedMore restrictive than Haiku 4.5, blocks pentesting; first Haiku with built-in cyber safeguards for a narrow high-risk set (FACT per Reuters)
SDK toolingStandard tool useBetaComputer use + browser use in beta in Python/TS SDKs
⚙

How it works

  • Tier architecture. Haiku is Anthropic's small-model tier inside a family where Opus 5.5 handles sustained judgment work, Sonnet 5.5 handles everyday/agent work, and Haiku 5.5 is positioned as the high-volume, latency-sensitive worker (classification, extraction, routing, summaries, compaction, database queries, subagent lookups).
  • Effort setting. Like larger Claude models, Haiku 5.5 accepts an effort control (Low→Max); higher effort raises per-task cost and intelligence, letting developers tune cost/quality per request class. The launch page charts accuracy-vs-cost curves on OSWorld 2.1, GDPval-AA, and HLE across effort levels.
  • Two-tier prompt pricing. Requests with prompts ≤100,000 input tokens are billed at the low band; everything above (including cache reads/writes and batch) is billed at the higher band. Each request is priced on its own; a request crossing the threshold pays high-band prices even for cache hits. Around 90% of Haiku 4.5 requests fell in the low band (FACT cited by Anthropic).
  • New tokenizer. Haiku 5.5 uses the post-4.7 tokenizer (~30% more tokens than the older tokenizer for the same text, per Anthropic pricing docs) — this is why headline per-token cuts don't translate 1:1 into per-task savings, and why Anthropic's "75% average" figure is tokenizer-adjusted.
  • Context and modality. 1M-token context; text and image input with text output (per Anthropic models overview; multimodal per Emergent's coverage).
  • Safety architecture. Stronger alignment results claimed versus Haiku 4.5; cybersecurity safeguards more restrictive than Haiku 4.5's but somewhat less restrictive than Sonnet 5.5's — permitting a wider range of defensive cyber tasks while blocking penetration testing and attacker-typical techniques; biology safeguards equivalent to Sonnet 5/Sonnet 5.5/Opus 5, with Life Sciences and Cyber Verification Programs for elevated access. Reuters notes this is Anthropic's first Haiku with built-in safeguards covering a narrow set of high-risk cybersecurity requests.
  • Distribution. First-party Claude API plus Bedrock, Vertex AI, and Foundry; regional/multi-region endpoints and data-residency multipliers (1.1x for US-pinned inference) apply as with other Claude 4.6+ models.
!

Why it matters

🎓 For Explorer
  • Unit economics of agentic pipelines. Agent loops multiply token counts: a single task can trigger dozens of model calls. Cutting the workhorse-model price by ~75% on average changes the break-even point for always-on agents (support triage, monitoring loops, RAG re-scoring, subagent fan-out), making workloads viable that were not viable at $1/$5.
  • A direct pricing fight with OpenAI's GPT-6 Luna. Both models sit at $0.10/$0.50 (FACT: Anthropic pricing docs and OpenAI pricing as reported by Powerdrill, DataCamp, Simon Willison). Anthropic's published benchmark table shows Haiku 5.5 ahead of Luna on every row where both appear (COMPANY CLAIM); Artificial Analysis's AA-Briefcase (1578 vs 1336) and Vals AI independent runs support Haiku being at least competitive, though Luna keeps a cheaper cost-per-point profile at max effort (INDEPENDENT EVIDENCE via the Data Science Collective/Medium comparison).
  • Small-model capability inflation. Computer use near the human baseline (~72% on OSWorld, per TechTimes citing the original eval paper) at $0.10/M input is a step-change for UI-automation products, which previously had to run large, expensive models for the same task.
  • IPO-adjacent revenue motion. Reuters and Yahoo Finance contextualize the launch ahead of Anthropic's planned IPO (June confidential filing; expected before Thanksgiving). A high-volume, low-price tier drives API consumption and revenue scale, and the accompanying Max/Team API credits pull subscription users into platform usage.
✦

What became possible?

🎓 For Explorer
  • Subagent fan-out at scale: a large planner (Opus/Sonnet) delegating lookups, extraction and drafting passes to Haiku 5.5 at $0.10/$0.50 — the pattern already used by customers quoted on the launch page (e.g., Rogo's "bigger model builds the deck, Haiku subagent pulls the segment revenue line from the 10-K").
  • Always-on, latency-sensitive assistants: live customer support, voice agents, in-app assistants where each turn previously cost too much or took too long (Asana reports >30% latency reduction and up to 2.5x faster inference per agent turn; Box reports 11 points higher than Haiku 4.5 at about half the latency — both COMPANY CLAIM customer quotes).
  • Computer-use and browser-use automation at commodity prices, with the SDK beta tooling landing at the same time.
  • Batch-scale enrichment at $0.05/$0.25 per MTok via the Batch API (50% discount).
◎

Implications

Technical

  • Performance-per-dollar is now the battleground metric. The release defines the small-model tier by cost-per-task rather than per-token, because the tokenizer change means per-token comparisons mislead.
  • Effort dials become part of the system design. Developers now have a continuous cost/intelligence control on the cheap tier, shifting cost optimization from prompt engineering to effort allocation across multi-stage pipelines.
  • Long-context pricing is non-linear. Unique among current Claude models, Haiku 5.5 has a prompt-length price cliff at 100K tokens (5x input, 5x output). With a 1M context and a ~30% heavier tokenizer, long-context Haiku workloads can be dramatically more expensive than the headline suggests — a trap for naive migration of 200K-context Haiku 4.5 jobs.
  • Independent eval variance is real. Artificial Analysis measured Terminal-Bench 4.0 at 33% and flagged AutomationBench-AA at 35% as likely understated due to a pre-release safety refusal issue (over-refusal), with Anthropic working on a fix; Vals AI measured Terminal-Bench 4.0 at 35.35% in its own harness. Vendor (39.2%) and independent (33–35%) numbers are close but not identical — the gap matters for claims-grade evaluation.
  • Tokenizer discontinuity. New-tokenizer models produce ~30% more tokens for identical text, inflating raw token spend even where per-token prices fall; workload-level (not token-level) comparisons are required.

Developer

  • Migration path. Anthropic publishes a migration guide (https://platform.claude.com/docs/en/models/haiku-5-5/migration-guide) for moving from Haiku 4.5; model ID change (claude-haiku-5-5), tokenizer change, and effort-setting behavior are the three main deltas.
  • Cost math changes. Prompt caching is extremely cheap on the low band ($0.01 cache reads) — amortizing large system prompts across many calls is now nearly free. Batch API at $0.05/$0.25 makes offline enrichment jobs a rounding error.
  • Tool-use overhead fell. Tool-use system prompt token counts dropped from 496/588 (Haiku 4.5) to 286/406 (Haiku 5.5), trimming per-call overhead for agentic apps.
  • New beta tooling. Python/TypeScript SDK betas for computer use and browser use target exactly this model's speed/price profile; developers building UI agents should evaluate them now.
  • Evaluation discipline. Validate on cost-per-task and accuracy-per-task, not price-per-token; the effort setting materially changes both. Run your own eval suite per Anthropic's own advice (mirrored by Powerdrill, DataCamp, Kingy: test on your files).

Enterprise

  • High-volume workload economics. RAG re-scoring, document extraction, classification, ticket triage, and recurring report generation (Box's stated use) become sub-cent-per-item categories at scale.
  • Agent platform build-out. The Opus/Sonnet-planner + Haiku-worker pattern is now cheap enough for broad enterprise rollout, not just thin pilots; subagent orchestration becomes a default architecture rather than an optimization.
  • Voice and support automation. Latency at ~240 tokens/sec (max, Artificial Analysis) and low per-turn cost make live-support and voice-agent loops economically sound, a shift over the previous Haiku tier.
  • Procurement math. Enterprises negotiating annual API commitments face lower per-token costs but more tokens (new tokenizer) and the >100K price cliff — negotiation baselines should be per-task.
  • Cloud distribution parity. Availability on Bedrock, Vertex, and Foundry with the same per-token list price keeps hyperscaler-neutral procurement viable; US data-residency pinning costs 1.1x.

Strategic

  • The small-model price war is now symmetrical. OpenAI fired first with GPT-6 Luna at $0.10/$0.50 (Sept 22); Haiku 5.5 matches at parity while showing stronger vendor-reported benchmarks and credible independent evals. DeepSeek V4.1 Flash (~$0.15/$0.60, open weights) and other open-weight vendors keep the floor moving. (COMPANY CLAIM for Haiku's benchmark advantage; INDEPENDENT EVIDENCE partially supports it.)
  • Anthropic's cadence vs. its pacing rhetoric. The Claude 5.5 family shipped three models in ~2.5 weeks (Opus 9/22, Sonnet 9/28, Haiku 10/7 — FACT). Commentary (shattered.io) notes tension with Anthropic's public pacing-plea positioning; the quarterly revenue pressure ahead of an IPO is the counterweight (INTERPRETATION).
  • Platform economics over margin. Cutting the workhorse tier's price 75%+ signals Anthropic is choosing volume, ecosystem lock-in (SDK betas, API credits, cache pricing), and consumption growth over near-term per-token margin — the same playbook OpenAI used for Luna (INTERPRETATION).
  • Haiku is re-positioned as a subagent fabric, not a chat model. Anthropic's own framing (subagent to Opus/Sonnet; compaction; extraction) makes the small model part of a system, not a standalone product — tightening the multi-model dependency within the Claude platform (INTERPRETATION).
⚠

Risks & limitations

Risks
  • Self-reported benchmark inflation (COMPANY CLAIM vs INDEPENDENT EVIDENCE). Vendor numbers (e.g., OSWorld 72.4%, Terminal-Bench 39.2%) are directionally confirmed but not exactly reproduced (AA 33%, Vals 35.35% on Terminal-Bench); headline claims should be treated as marketing-grade until reproduced on your own workloads.
  • "Human-level" computer use is partly an artifact. TechTimes, citing Epoch AI's independent analysis of OSWorld, notes ~15% of OSWorld tasks can be completed with terminal shortcuts and ~30% with scripting instead of genuine GUI navigation — the 72.4% score overstates real GUI competence. The OSWorld result is on the 82-task offline subset with max effort.
  • Over-refusal in deployment. Artificial Analysis measured an AutomationBench over-refusal issue in pre-release testing (35% vs 53–60% for comparable models); a safety-refusal fix is pending, and until then agentive workloads may see spurious refusals.
  • Long-context cost cliff. 1M context invites long-prompts; a request over 100K input tokens jumps to 5x prices including on cache reads (per pricing docs), so naive long-document migrations can surprise finance.
  • Effort-setting quality variance. Default (medium) effort scores materially below max (Kingy cites system-card medium-effort GDPval-AA 1277 vs 1620 max; Briefcase 1372 vs 1578) — teams that forget to tune effort leave quality on the table or pay for unused depth.
  • Distillation/IP pressure. Cheaper, highly capable small models reduce the cost of distilling proprietary behaviour; Anthropic's preserved-thinking and distillation-defense research (S12) is the countermeasure (context).
Limitations
  • All headline benchmark comparisons are COMPANY CLAIMS from Anthropic's launch page and system card, run at max effort under Anthropic's evaluation protocol; independent runs (Artificial Analysis, Vals AI) confirm direction but differ in magnitude and harness.
  • Customer quotes (Asana, Box, HubSpot, AlphaSense, Rogo, Cognition) are anthropic-selected early-testing testimonials — COMPANY CLAIM, not independent evaluations.
  • The "~75% lower average cost" figure is Anthropic's own tokenizer-adjusted calculation; per-token savings are 90% (<100K) and 50% (>100K), and the new tokenizer partially offsets per-task savings.
  • OSWorld and Terminal-Bench results are single-configuration measurements (max effort, offline subset for OSWorld); production performance will vary with effort, prompting, and tooling.
  • Not all claims were independently reproduced at research time; no third-party cost-per-task audit (outside Artificial Analysis' index economics and Vals AI's cost/test figures) existed within the window.
?

Open questions

  • Will Artificial Analysis's promised AutomationBench re-run (post over-refusal fix) land closer to vendor numbers, and does the fix ship in this window?
  • How fast do OpenAI (Luna price moves / effort features), Google (Gemini 4 Argon — not yet developer-open) and open-weight vendors (DeepSeek V4.1 Flash) respond to Haiku 5.5's price/quality point?
  • What share of Haiku 4.5 traffic actually migrates, given tokenizer and effort differences?
  • Does the >100K price cliff push long-context users to Sonnet 5.5 ($2/$10, flat) instead?
  • How much of the price cut is architecture (efficiency) vs. margin strategy ahead of the IPO?
  • Will the new monthly Max/Team API credits measurably shift consumer subscription usage onto the API?
↗

What happens next?

🎓 For Explorer

PREDICTION: (1) Anthropic pushes the over-refusal fix and re-publishes AutomationBench results within weeks; (2) OpenAI responds on Luna's long-context band (currently 272K before 2x/1.5x rises vs Haiku's 100K cliff) and/or pricing; (3) Google opens Gemini 4 Argon to developers, completing three-way small-model competition; (4) migration waves from Haiku 4.5 follow, but tokenizer/effort surprises will produce visible "cost-per-task actually went up" anecdotes in the long-context band; (5) Anthropic's IPO filing season (before Thanksgiving per Yahoo Finance) will lean on API consumption growth stories anchored by this price/volume trade; (6) expect further small-model price cuts from open-weight vendors, keeping margin pressure on the closed small tiers into Q1 2027.

★

Editorial takeaway

🎓 For Explorer

The small-model tier just became the most contested battleground in AI — and intelligence-per-dollar, not raw IQ, is the unit of war. Haiku 5.5's significance is not that a small Anthropic model got better; it is that Anthropic matched OpenAI's price point exactly while posting (vendor-reported, partially independently confirmed) quality that outperforms the rival at that price, and wrapped it in the operational machinery — effort dials, sub-cent cache reads, subagent positioning, SDK betas, and API credits — that turns a price cut into a platform play. The sceptical reading is equally important: headline savings are tokenizer-adjusted, the "human-level computer use" headline carries Epoch AI caveats, an over-refusal bug shipped, and the 100K-token cliff is a genuine footgun. Enterprises should adopt the economics fast, but validate on their own workloads — because in a war measured in tenths of a cent, the difference between a marketing table and your production telemetry is the whole game.

Lab: COMPARE

Research sources

Primary Sources (3)
Primary
Anthropic — Claude Haiku 5.5 System Card (companion to launch page; PDF) URL: https://www.anthropic.com/claude-haiku-5-5-system-card Published: 2026-10-07 Supports: Evaluation methodology for the launch-page benchmarks; referenced via the launch page and independently cited by kingy.ai (medium-effort GDPval-AA 1277 / Briefcase 1372; OSWorld 2.1 eval config: 82 offline tasks, max effort; Artificial Analysis attribution of GDPval/Briefcase runs). Direct PDF parsing was not possible in this environment; contents corroborated via the launch page and secondary cites. Evidence role: Primary methodology document (referenced); figures COMPANY CLAIM.
URL unavailable
Primary
Anthropic — Claude Platform pricing documentation URL: https://platform.claude.com/docs/en/about-claude/pricing Checked: 2026-10-13 (live during research) Supports: Standard -- Haiku 5.5 list pricing (both bands), cache write/read pricing ($0.125/$0.625 writes; $0.01/$0.05 hits), Haiku 4.5 comparison ($1/$5, $1.25 writes, $0.10 hits), Batch API rates (Haiku 5.5 $0.05/$0.25 <100K), long-context pricing mechanics (per-request banding, cache reads counted in prompt length), new tokenizer note (~30% more tokens), tool-use system-prompt token counts (286/406 vs Haiku 4.5's 496/588), data-residency 1.1x multiplier. Evidence role: Primary documentation; pricing facts are FACT (first-party list prices).
URL unavailable
Primary
Anthropic — "Claude Haiku 5.5" (official launch page) URL: https://www.anthropic.com/claude-haiku-5-5 Published: 2026-10-07 Supports: Announcement, model ID `claude-haiku-5-5`, "cheapest, fastest, most capable small model" positioning, ~75% average cost reduction vs Haiku 4.5 (footnote 2: 90% lower <100K / 50% >100K, tokenizer-adjusted), full pricing table ($0.10/$0.50 within 100K band; $0.50/$2.50 above; cache $0.01/$0.05 reads, $0.125/$0.625 writes), benchmark table (GDPval-AA 1620, AA-Briefcase 1578, OSWorld 2.1 72.4%, HLE 45.9%/57.4%, Terminal-Bench 4.0 39.2%, FrontierCode 46.4%, Chartography 46.4%), first Haiku with adjustable effort, safety section (cyber safeguards, biology safeguards, verification programs), availability (AWS/GCP/Azure), Sonnet 5.5 cache-read price cut to $0.10, Max/Team API credits, SDK beta computer/browser use, customer quotes (Asana, Box, HubSpot, AlphaSense, Rogo, Cognition). Evidence role: Primary announcement; all capability/benchmark figures are COMPANY CLAIM unless independently corroborated.
URL unavailable
Independent Sources (6)
Independent
Simon Willison — "Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war" URL: https://simonwillison.net/2026/Sep/22/opus-and-sol-and-luna Published: 2026-09-22 Supports: Pre-launch context: Haiku 4.5 at $1/$5 vs GPT-6 Luna at $0.10/$0.50; the price war at the tier below flagship; GPT-6 Luna pricing table (in/cached/out); expectation that Haiku needed to regain price competitiveness. Evidence role: Independent expert context for the competitive price-gap the release closed.
URL unavailable
Independent
Yahoo Finance — "Anthropic reveals Haiku 5.5 model as AI pricing war intensifies" URL: https://finance.yahoo.com/technology/article/anthropic-reveals-haiku-55-model-as-ai-pricing-war-intensifies-180000423.html Published: 2026-10-07 Supports: "AI pricing war" framing; 75% average cost claim; $0.10/$0.50 pricing "in line with" OpenAI's GPT-6 Luna; Anthropic IPO expected before Thanksgiving; open-weight alternatives pressure. (Accessed via search-result excerpts.) Evidence role: Independent business reporting on pricing strategy and IPO context.
URL unavailable
Independent
TechTimes — "Claude Haiku 5.5 Becomes Viable Sub-Agent: Human-Level Computer Use at Ninety Percent Lower Cost" URL: https://www.techtimes.com/articles/328793/20261009/claude-haiku-55-becomes-viable-sub-agent-human-level-computer-use-ninety-percent-lower-cost.htm Published: 2026-10-09 Supports: OSWorld 2.1 72.4% matching the ~72% human baseline from the original eval paper; GDPval-AA and AA-Briefcase comparisons (1620 vs 735 vs 1437; 1578 vs 1336 vs 614); Terminal-Bench 4.0 gap (39.2% vs 0% vs 16.4%); Epoch AI caveats that ~15% of OSWorld tasks are completable via terminal shortcuts and ~30% via scripting rather than GUI navigation. Evidence role: Independent reporting adding methodological caveats to headline claims.
URL unavailable
Independent
Vals AI — "Claude Haiku 5.5" model page (independent benchmark harness) URL: https://www.vals.ai/models/anthropic_claude-haiku-5-5 Published: 2026-10-07 (release data) — page live during research Supports: Independent Vals Index 54.31% (rank 16/45), Terminal-Bench 4.0 35.35% (rank 12/45) in its own harness, Vibe Code Bench v1.1 90.44% (rank 3/110), Harvey Legal Agent Benchmark, cost/test $2.991, refusal rate 0.22%, context 1M, max output 128K, in/out $0.10/$0.50. Evidence role: Independent benchmark corroboration (direction and rough magnitude vs vendor numbers).
URL unavailable
Independent
Artificial Analysis — "Anthropic has released Claude Haiku 5.5" (independent evaluation) URL: https://artificialanalysis.ai/articles/claude-haiku-5-5 Published: ~2026-10-07/08 Supports: Independent Intelligence Index 43 (max) vs Sonnet 5.5 56; output speed ~240 t/s (max) fastest in family; AA-Briefcase v1.1 Elo 1578 (ahead of Kimi K3, GLM-5.3; comparable to Muse Spark 1.3 max; within CI of GPT-6 Astra max and Fable 5.1 high); Terminal-Bench 4.0 33% vs 0% for Haiku 4.5, level with GLM-5.3 Flash, ahead of Gemini 3.8 Flash (20%) and GPT-6 Luna (13%); AutomationBench-AA 35% flagged as understated due to pre-release over-refusal, re-run expected; 1M context (up from 200K). Evidence role: Independent verification of capability direction; specific scores differ from vendor claims.
URL unavailable
Independent
Reuters — "Anthropic launches third Claude 5.5 model, expanding AI lineup before planned IPO" URL: https://www.reuters.com/business/anthropic-launches-third-claude-55-model-expanding-ai-lineup-before-planned-ipo-2026-10-07 Published: 2026-10-07 Supports: Independent confirmation of launch date, third-in-family status, "costs 75% less than its predecessor," $0.10/$0.50 and $0.50/$2.50 pricing tiers, target workloads (classification, summarization, extraction, live support, voice agents, in-app assistants), first Haiku with built-in safeguards for a narrow set of high-risk cybersecurity requests, IPO context. (Article body accessed via search-result excerpts; direct fetch blocked by paywall.) Evidence role: Independent news confirmation of the event and headline pricing.
URL unavailable
Secondary Sources (8)
Secondary
Quartz (via official Facebook post) — "Anthropic cuts Claude Haiku 5.5 prices by up to 90%" URL: https://www.facebook.com/quartznews/posts/icymi-anthropic-is-cutting-prices-on-its-smallest-ai-model-by-up-to-90-the-new-c/1461207279208419 Published: ~2026-10-08 Supports: Independent mainstream summary: price cuts up to 90%, $0.10/M input, price parity with GPT-6 Luna. Evidence role: Secondary corroboration of the pricing headline.
URL unavailable
Secondary
MLQ News — "Anthropic Launches Claude Haiku 5.5 With 90% Lower API Prices" URL: https://mlq.ai/news/anthropic-launches-claude-haiku-55-with-90-lower-api-prices Published: 2026-10-07 Supports: Up-to-90%-lower per-token pricing framing, computer-use improvement, high-volume workload positioning (classification, summarization, database queries, customer support, subagent tasks), availability incl. AWS/Google Cloud/Azure, SDK beta computer/browser use support, caveat that benchmark/price comparisons depend on workload, effort and token use. (Direct fetch returned 403; content accessed via search-result excerpts.) Evidence role: Secondary AI-news corroboration of pricing and availability.
URL unavailable
Secondary
Technology.org — "Claude Haiku 5.5 Cuts Token Prices by Up to 90%" URL: https://www.technology.org/2026/10/08/claude-haiku-5-5-anthropic-price-benchmarks Published: 2026-10-08 Supports: Two-tier pricing summary ($0.10/$0.50 up to 100K; $0.50/$2.50 above; cache $0.01/$0.05 and $0.125/$0.625), ~75% average claim, OSWorld 72.4% / Terminal-Bench 39.2% figures, third-model-in-two-weeks framing, June confidential IPO filing context. Evidence role: Secondary summary corroboration.
URL unavailable
Secondary
Activepieces blog — "Claude Haiku 5.5 Cost and Performance Benchmarks" URL: https://www.activepieces.com/blog/claude-haiku-55-cost-and-performance-benchmarks Published: 2026-10-09 Supports: Independent hands-on signal: invoice extraction in 1.0s via OpenRouter on 2026-10-09; 980/1000 invoices automated vs GPT-4o-mini hallucinating currency symbols; "fastest model in the lineup" framing; pricing cross-checked against anthropic.com, platform.claude.com, openrouter.ai on Oct 9, 2026. Evidence role: Secondary practical performance/cost datapoint.
URL unavailable
Secondary
Kingy.ai blog — "Claude Haiku 5.5: Benchmarks, Specs, Sol & Luna Compared" URL: https://kingy.ai/blog/claude-haiku-5-5-benchmarks-specs-sol-luna Published: ~2026-10-08 Supports: System-card detail (medium-effort GDPval-AA 1277 / Briefcase 1372 vs max; eval config for OSWorld 2.1: 82 offline tasks, 5 attempts, 1080p, max effort), expanded OSWorld table incl. GPT-6.1 Sol (76.6%) and Opus 5.5 (87.2%), briefcase-caveat discussion. Evidence role: Secondary source corroborating system-card specifics.
URL unavailable
Secondary
DataCamp blog — "Claude Haiku 5.5 vs GPT-6 Luna: Benchmarks, Pricing, and a Hands-On Test" URL: https://www.datacamp.com/blog/claude-haiku-5-5-vs-gpt-6-luna Published: ~2026-10-08/09 Supports: Independent hands-on comparison: same list prices ($0.10/$0.50, cache reads $0.01), Haiku 5.5 ahead of Luna on every Anthropic-listed benchmark row, long-context band differences (Haiku rises above 100K; Luna only above 272K at 2x input/1.5x output), practical cost observation of Luna being cheaper in test at equal prices (tokenizer/band effects). Evidence role: Secondary hands-on comparison reinforcing band/tokenizer cost caveats.
URL unavailable
Secondary
Powerdrill Bloom blog — "Claude Sonnet 5.5 and Haiku 5.5: What's New, Pricing, and Alternatives (2026)" URL: https://powerdrill.ai/blog/claude-sonnet-5-5-and-haiku-5-5 Published: 2026-10-09 Supports: Cross-checks family dates (Opus 9/22, Sonnet 9/28, Haiku 10/7 via three official channels), pricing table incl. cache bands, context/output limits (1M / 128K), effort-setting guidance, alternatives table (GPT-6 Luna $0.10/$0.50 272K context, GPT-6.1 Sol $2/$10, Gemini 4 Argon not yet open), customer-quote analysis (Box, Rogo), "vendor pages read Oct 8, 2026" sourcing. Evidence role: Secondary pricing and availability corroboration.
URL unavailable
Secondary
Emergent.sh News — "Anthropic Launches Claude Haiku 5.5 Multimodal Model" URL: https://emergent.sh/news/anthropic-launches-claude-haiku-55-multimodal Published: 2026-10-08 Supports: Confirms launch date (Oct 7, 2026), availability via API and partner platforms, multimodal (text + vision) positioning, family composition (Opus 5.5, Sonnet 5.5, Haiku 5.5, Fable 5.1), proprietary licensing. Evidence role: Secondary news confirmation of event and availability.
URL unavailable
Up next · story 4 of 20
Mistral's 'le Chonk' hits the top tier outside the US and China, but its weights are not out yet
All storiesNext story