Claude Sonnet 5.5: Pricing, Benchmarks, and Tested Performance
Claude Sonnet 5.5 is here at $2/$10 per million tokens. Tested across 13 benchmark tasks, compared with Opus 5.5 and GPT-6.1 Sol, plus Haiku 5.5 details.
Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.
👁 118 reads
Regularly updated. As of September 29, 2026, including my multi-task benchmark evaluation.
Claude Sonnet 5.5 launched on September 28, 2026, holding the same API pricing as Sonnet 5: $2 per million input tokens and $10 per million output tokens. One day after launch, I ran it through my evaluation suite across 13 tasks tested in Polish: it scored 80.4 out of 100, up 8.5 points from Sonnet 5, but 9.7 points behind Claude Opus 5.5. It proved to be the fastest among five Anthropic models in my tests, processing the entire suite for $0.095. Below you will find the scores, sample outputs, comparisons with Opus 5.5 and GPT-6.1 Sol, and the breakdown of Claude Haiku 5.5. Haiku 5.5 arrived on October 7, 2026 - pricing, access, and benchmark numbers are detailed in my Claude Haiku 5.5 guide.
Overview
Sonnet 5.5 is available directly in the Claude app, via the API under claude-sonnet-5-5, and through Amazon Bedrock, Google Cloud, and Microsoft Foundry. In Anthropic’s internal evaluations, it matches Opus 5.5 in knowledge work (GDPval-AA 1844 vs 1846) and pulls ahead in Terminal-Bench 4.0 (70.6% vs 66.4%), at half the price. In my non-English writing evaluation (tested in Polish), it trailed Opus 5.5 noticeably (80.4 vs 90.1 points) and lagged behind GPT-6.1 Sol (85.8 points), which sells at identical pricing. However, it completed the 13 tasks in just 83 seconds, compared to 188 seconds for Opus 5.5 and 140 seconds for Sonnet 5. For polished long-form copy destined for publication, Opus remains the superior choice; for emails, translations, proofreading, and code, Sonnet 5.5 delivers compelling efficiency. Haiku 5.5 launched on October 7, 2026, at $0.10 / $0.50 and posted 76.8 points on the same test suite.
What Anthropic Officially Confirmed
Every detail in this section comes directly from the Sonnet 5.5 release documentation, API pricing sheets, and developer guides verified on September 29, 2026. Sonnet 5.5 joins Opus 5.5 (September 22) and Fable 5.1 (September 1) as the third model featuring a 1-million-token context window with output capacity reaching up to 128k tokens.
- Pricing: $2 per million input tokens and $10 per million output tokens, prompt cache reads at $0.20, cache writes at $2.50, and batch mode at $1 / $5.
- Speed and cost: Anthropic claims over 30% faster token generation and up to a 30% reduction in total task cost compared to Sonnet 5, as the model requires fewer tokens to complete requests.
- Reasoning effort: Defaults to “medium” inside the Claude user interface and Claude Code, and “high” in direct API calls.
- Safety and security: The first Sonnet model equipped with dedicated cybersecurity guardrails; flagged queries automatically fall back to Sonnet 5.
- Knowledge cutoff: Firmly established at June 2026.
- Developer changes: Anthropic notes 5 breaking changes relative to Sonnet 5, including setting reasoning effort to between_tools instead of disabled to turn off intermediate thinking, along with removed forced tool calls. These are detailed in “What’s new in Sonnet 5.5”.
- Haiku 5.5: Released October 7, 2026, priced at $0.10 / $0.50 per million tokens (20 times cheaper than Sonnet 5.5), sporting a full 1-million-token window. Complete analysis is in my guide to Claude Haiku 5.5.
Pre-Launch Leaks vs Reality
- Leak - “Lyra” on X (Sept 24 and 26): Correctly anticipated the $2 / $10 pricing and a release “next week”. Confirmed.
- Leak - Factory Droid 0.228.0 (Sept 27): Exposed the claude-sonnet-5-5 identifier behind a feature flag. Confirmed.
- Unconfirmed: Whether Sonnet 5.5 serves as the default engine on Claude’s free tier. The announcement notes availability “across all platforms”, while pricing tiers list Sonnet across every plan without an explicit version tag.
- Haiku 5.5 rumor: Speculated at roughly 1/20th of Opus pricing. Proven inaccurate - Haiku 5.5 arrived at 1/40th of Opus 5.5’s cost.
Hands-On Evaluation and Benchmark Scores
I evaluated Sonnet 5.5 via OpenRouter (anthropic/claude-sonnet-5.5) against the identical 13 prompts used in my ongoing testing suite (we tested it in Polish): 9 creative and functional writing prompts (article, product copy, client email, LinkedIn post, short story, SEO headlines, summary, explainer for a child, translation), two proofreading challenges, and two complex linguistic quizzes. Settings mirrored baseline tests: reasoning effort set to “low”, standardized token ceilings. The overall score assigns 70% to blind judge scoring, 20% to strict adherence to constraints, and 10% to clean formatting (avoiding excessive em-dashes and filler prose).
Outputs were evaluated in a blind protocol by two independent frontier models: GPT-6 Sol and Gemini 3.1 Pro. Anthropic’s own Claude Opus 5.5 was excluded from the judging panel to avoid self-evaluator bias. Alongside Sonnet, the judges graded four baseline anchor models from the initial run (Opus 5.5, GPT-6 Luna, Qwen3.8 Max, GLM-5.3). Relative rankings held consistent; without the Anthropic judge, three of the four anchors received slightly higher marks. The overall calibration drifted upward by an average of 1.3 points, so Sonnet’s final score was adjusted downward by that exact delta. Anchor variance between runs remained tight at 0 to 2.2 points.
| Model | Score | Correctness | Naturalness | Execution | Hard Proofreading | Hard Quiz |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 90.1 | 9.89 | 8.50 | 8.56 | 30/30 | 48/48 |
| GPT-6.1 Sol | 85.8 | 9.50 | 7.39 | 8.28 | 28/30 | 48/48 |
| GPT-6 Astra | 85.3 | 9.56 | 7.56 | 8.44 | 29/30 | 48/48 |
| GPT-6 Luna | 82.8 | 9.50 | 7.50 | 7.61 | 29/30 | 48/48 |
| Claude Sonnet 5.5 | 80.4 | 9.33 | 7.78 | 7.72 | 29/30 | 47/48 |
| Claude Sonnet 5 | 71.9 | 8.44 | 6.56 | 6.50 | 23/30 | 47/48 |
Judges evaluated criteria on a 1-10 scale. Sonnet 5.5: September 29, 2026 run, calibrated via baseline anchors; GPT-6.1 Sol: same-day run; remaining models from the initial run. Sonnet 5.5 achieved a perfect score on both the standard proofreading exercise (22/22) and the standard quiz (46/46).
Cost, Latency, and API Pricing
| Model | 13-Task Suite Cost | Total Latency | Output Tokens / Sec | API Pricing per 1M Tokens (Input / Output) |
|---|---|---|---|---|
| Claude Opus 5.5 | $0.372 | 188 s | 92 | $4 / $20 |
| GPT-6 Astra | $0.337 | 146 s | 40 | $10 / $50 |
| GPT-6.1 Sol | $0.073 | 110 s | 58 | $2 / $10 |
| Claude Sonnet 5.5 | $0.095 | 83 s | 99 | $2 / $10 |
| Claude Sonnet 5 | $0.102 | 140 s | 64 | $2 / $10 |
Costs reflect OpenRouter invoicing; latency includes end-to-end network round trips; provider pricing excludes local sales tax. Because tokenizers vary across providers, output tokens per second should only be directly compared within the same model family.
The speed advantage is clear: Sonnet 5.5 finished the full prompt suite 41% faster than Sonnet 5 and generated 54% more output tokens per second. The cost reduction was more moderate: $0.095 vs $0.102, a 7% drop rather than Anthropic’s stated “up to 30%”. Anthropic’s benchmark savings stem primarily from large coding workloads, whereas concise non-English prompts show smaller margins. Opus 5.5 cost nearly 4 times more on this exact test despite its base list price being only double. The reason: on “low” reasoning, Opus 5.5 allocated 8,867 tokens to internal reasoning traces, while Sonnet 5.5 allocated none. OpenRouter billing confirms Sonnet 5.5 bypasses extended reasoning altogether under the low setting.
Output Excerpts and Observations
Sonnet 5.5 shone most brightly on functional tasks: the client communication email (9.33), the English-to-Polish translation task (9.0), and creative narrative fiction (8.83). In a short story depicting a retired mail carrier delivering letters to former apartment tenants, it closed with the following scene (translated from the Polish test output):
“This morning, an envelope heavier than the rest sat in the mail slot. The paper was cream, the penmanship neat, sloping, unmistakably her own. Recipient: ‘Halina Wrobel, 7th floor, door 43.’ Sender: ‘Halina Wrobel, some time from now.'”
Gemini 3.1 Pro commended the excerpt for natural, literary texture devoid of artificial filler. The opening hook for a small ceramic studio’s LinkedIn profile also succeeded: “In my first month, I sold exactly one mug. My mother bought it.” GPT-6 Sol noted, however, that the phrasing carried slight grammatical rigidity in the original language.
Weak spots were evident. In an English-to-Polish newsletter translation where literal idioms were explicitly banned, the model left an unadapted calque of “the ball is in our court”. When explaining inflation to an eight-year-old, it substituted an awkward technical phrase instead of standard purchasing power phrasing. In a household energy guide, it slipped into colloquial regional phrasing and generated generic, templated section headers.
Strengths and Weaknesses
- Strengths: Concise functional copy (email 9.33, translation 9.0, child explainer 8.67), proofreading accuracy (22/22 and 29/30 corrections found, up from Sonnet 5’s 23/30), broad regional factual recall (93 of 94 questions correct), high throughput, and minimal latency.
- Strengths: Noticeable improvement over Sonnet 5 across all judge metrics: correctness rose from 8.44 to 9.33, naturalness from 6.56 to 7.78, and structural execution from 6.50 to 7.72.
- Weaknesses: Loose adherence to strict length boundaries. A product description paragraph came in at 90 words against an explicit 120-150 word instruction (earning a 7.0), while a summary reached 85 words against an 80-word ceiling (7.83).
- Weaknesses: Susceptible to trick historical questions with false premises regarding early 20th-century electoral mechanics.
- Weaknesses: High frequency of long dashes (9.4 per 1,000 words vs 5.1 for Opus 5.5) and repetitive, cookie-cutter subheaders.
Sonnet 5.5 vs Opus 5.5: Decision Guide
| Use Case | Recommended Model | Key Rationale |
|---|---|---|
| High-stakes published text: long articles, essays, fiction | Opus 5.5 | 90.1 vs 80.4 in benchmark testing; largest gap in instruction execution (8.56 vs 7.72) and tight length compliance |
| Emails, localization, copyediting, rapid iterations | Sonnet 5.5 | 9.0-9.33 scores, proofreading rates of 22/22 and 29/30, completed in under half the time |
| Agentic coding and command-line execution | Sonnet 5.5 daily, Opus 5.5 for architectural refactoring | Anthropic data shows Terminal-Bench at 70.6% vs 66.4%, though FrontierCode favors Opus at 54.4% vs 52.1% |
| Complex, open-ended strategic reasoning | Opus 5.5 | Explicitly recommended by Anthropic for deep evaluation workflows |
| High-volume non-English text via API | Neither: evaluate alternatives | GPT-6.1 Sol reached 85.8 points at identical pricing, while GPT-6 Luna hit 82.8 points at a fraction of the cost |
| Claude Free Tier | Sonnet | Claude pricing reserves Opus exclusively for paid Pro and Team subscribers |
My takeaway: Sonnet 5.5 represents a substantial upgrade over Sonnet 5 and serves as an exceptional everyday workhorse, but it does not replace Opus 5.5 for high-craft editorial work destined for public release. Processing the 13-task test suite cost $0.372 on Opus and $0.095 on Sonnet; across modest daily volume, the absolute cost difference is negligible compared to the quality difference. Note on testing conditions: these benchmarks ran at the “low” reasoning tier, where Sonnet 5.5 engaged no extended thinking. The Claude web app defaults to “medium”, which may yield higher quality at the cost of additional output tokens.
Release Timeline: Sonnet 5.5 on September 28, Haiku 5.5 on October 7
Sonnet 5.5 launched September 28, 2026, six days after Opus 5.5. The Polymarket contract tracking “Next Claude Sonnet released on…?” resolved to September 28. For Haiku, prediction markets on September 29, 2026, showed the following odds:
| Market (Polymarket) | Implied Probability |
|---|---|
| New Claude Haiku by Oct 15, 2026 | 94% |
| New Claude Haiku by Oct 31, 2026 | 95% |
| New Claude Haiku by Dec 31, 2026 | 98% |
| Next Sonnet (5.6+) by Dec 31, 2026 | 29.5% (minimal volume) |
The prediction markets proved accurate: Haiku 5.5 landed on October 7, 2026, nine days following Sonnet 5.5. The preceding Haiku 4.5 model had launched October 15, 2025, while Haiku 5 was skipped entirely. Haiku 4.5 remains available in the API. You can follow upcoming releases in our AI release tracker.
Anthropic Benchmark Results and Haiku Baseline
Below are the official vendor figures reported on Anthropic’s release page. Note that Sonnet 5’s Terminal-Bench 4.0 score (10.3%) appears anomalously low, and independent auditor Artificial Analysis reported Sonnet 5.5 at 63.6% rather than 70.6% on that benchmark. Anthropic explicitly concedes that Opus 5.5 remains decisively superior on extended, open-ended evaluations requiring continuous contextual judgment.
| Benchmark | Sonnet 5.5 | Opus 5.5 | Sonnet 5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 66.4% | 10.3% |
| FrontierCode 1.1 | 52.1% (xhigh), 46.2% (max) | 54.4% | 42.4% |
| CursorBench 4.0 | 55.5% | 57.8% | 34.1% |
| GDPval-AA v2.1 (Elo) | 1844 | 1846 | 1449 |
| Humanity’s Last Exam (with tools) | 64.5% | 67.7% | 54.9% |
| OSWorld 2.1 | 80.1% | 81.8% | 57.0% |
On FrontierCode, Sonnet 5.5 performed worse at its maximum reasoning tier than at the tier below it. According to Anthropic, running at “max” triggered multi-agent code review loops that introduced out-of-scope edits, which penalize overall scores in this harness.
The 56-point index score for Sonnet 5.5 requires the maximum reasoning configuration, generating an intensive ~193,000 output tokens per task. At medium effort, it scores 41. Haiku 4.5 scored 17. Following its release, Haiku 5.5 registered 43 points (at max) - outscoring Sonnet 5 (38) while costing 10 times less than Haiku 4.5.
API Pricing and Plan Structure
Anthropic API price card verified September 29, 2026; net prices excluding local sales tax.
| Model | Input / 1M Tokens | Output / 1M Tokens |
|---|---|---|
| Claude Sonnet 5.5 | $2 | $10 |
| Claude Sonnet 5.5 (Batch Mode) | $1 | $5 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Haiku 4.5 (Previous Gen) | $1 | $5 |
Watch out for outdated references: certain sites still list Sonnet at $3 / $15. The proposed price increase originally planned for September 1 was abandoned, leaving Sonnet 5.5 at standard $2 / $10 rates. In Claude’s subscription app, the Pro tier runs $20 per month (or $17 monthly billed annually), while the Max tier starts at $100. Sonnet is accessible across all plans, including the free tier, while Opus requires Pro or higher. Global availability details are covered in our guide on Claude subscription plans and costs.
Haiku 5.5 Benchmark Verification
I ran Haiku 5.5 through the same 13-task test suite on October 8: it reached 76.8 points, compared to 80.4 for Sonnet 5.5, 59.2 for Haiku 4.5, and 82.8 for GPT-6 Luna ($0.004 total suite cost). A complete cost-efficiency breakdown is in Haiku 5.5 vs GPT-6 Luna vs Gemini 3.8 Flash, while practical guidance on swapping Haiku into developer setups is detailed in Haiku 5.5 in Claude Code. For broader model comparisons, see our dedicated reviews of Claude Opus 5.5, GPT-6.1 Sol, and GPT-6 Astra.
Frequently Asked Questions
When was Claude Sonnet 5.5 released?
September 28, 2026. The model is accessible via Claude web and desktop apps, directly in the API under claude-sonnet-5-5, and through Amazon Bedrock, Google Cloud, and Microsoft Foundry.
How much does Claude Sonnet 5.5 cost?
API pricing is $2 per million input tokens and $10 per million output tokens. This matches Sonnet 5’s original rate. Running my 13-task test suite cost $0.095 in total API usage.
How does Claude Sonnet 5.5 perform on complex text tasks?
In our benchmark suite (tested in Polish), it scored 80.4 out of 100: an 8.5-point gain over Sonnet 5 (71.9), but trailing Opus 5.5 (90.1) and GPT-6.1 Sol (85.8). It excelled at error correction and technical quizzes, but drifted on strict word-count limits.
Should you choose Sonnet 5.5 or Opus 5.5?
Use Opus 5.5 for publication-ready editorial prose, nuanced long-form content, and deep architectural assessments. Choose Sonnet 5.5 for everyday programming, technical workflows, translation, and high-frequency operational tasks where it cuts costs by 50% and doubles generation speed.
When did Claude Haiku 5.5 launch?
October 7, 2026, nine days following Sonnet 5.5. It lists at $0.10 / $0.50 per million tokens and scored 76.8 points in our test suite - full analysis is available in the Claude Haiku 5.5 overview.
Sources and Verification Date
Official documentation
- Anthropic, Claude Sonnet 5.5 (2026-09-28)
- Anthropic, Claude Opus 5.5 (2026-09-22)
- What’s new in Sonnet 5.5
- Model Overview
- API Pricing
- Claude Subscription Plans
- Supported Countries. Industry reporting: TechCrunch (2026-09-28)
- TestingCatalog. Data sources: Artificial Analysis on Sonnet 5.5
- Polymarket Haiku Tracker
- model specifications via OpenRouter. Evaluation testing: Evaluation suite
- September 29
- 2026 run
- 13 prompt tasks
- evaluated blindly by GPT-6 Sol and Gemini 3.1 Pro. Verified September 29
- 2026
- Haiku 5.5 data updated October 8
- 2026
More on this from QUANTSLANT
benchmarks·Claude Opus 5.5 Pricing: How Much Does It Cost?Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. We tested why a 20% price cut delivers 40% savings on real tasks.