Claude Sonnet 5.5 Review: 70.6% on Terminal-Bench at $2/$10

Anthropic released Claude Sonnet 5.5 on September 28, 2026, and it is the strangest kind of upgrade. The price did not move at all, yet the model behind it changed almost beyond recognition. It still costs $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. On the benchmark Anthropic cares about most for coding agents, Terminal-Bench 4.0, it scores 70.6%, up from 10.3% — and ahead of Opus 5.5, the flagship that costs twice as much per token. The catch is not in the price list. It is in the effort dial: run Sonnet 5.5 at low or medium effort and it is the best value in Anthropic's lineup, but crank it to Max and it can cost more per task than Opus 5.5 while scoring worse. This review covers the benchmarks, the full pricing with the fine print, and the one setting to avoid.

What Sonnet 5.5 Is

Sonnet 5.5 is the second model in the Claude 5.5 family, arriving six days after Opus 5.5. It is the mid-tier model in the range: below Opus 5.5, above Haiku 5.5, with the top of the range held by Fable 5.1 at $10/$50. Anthropic describes it as "the best combination of speed and intelligence" and pitches it as a faster, lower-cost complement to Opus, strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.

The model ID for developers is claude-sonnet-5-5. It is available on the Claude Platform, claude.ai, Claude Code, and Claude Cowork, and on all major cloud platforms: Amazon Web Services, Google Cloud, and Microsoft Azure. Artificial Analysis lists nine API providers carrying it. The context window is 1 million tokens at standard pricing, with up to 128,000 output tokens (300,000 on the Batch API with a beta header) and a June 2026 knowledge cutoff. Like Opus 5.5 and Sonnet 5, it is available with zero data retention.

One detail matters if you are migrating from Sonnet 5: the model uses adaptive thinking, and the lowest thinking setting is now between_tools, which turns off up-front thinking. Requests that set thinking: disabled return a 400 error, as do requests that set temperature, top_p, or top_k. The minimum cacheable prompt is 512 tokens, down from 1,024 on Sonnet 5, which makes it easier for agents with short system prompts to get cache hits.

The Benchmarks: Where It Beats Opus, and Where It Does Not

Every score below comes from Anthropic's launch announcement, so read them as the company's best case rather than an independent verdict. The independent check available so far is Artificial Analysis, which scores Sonnet 5.5 at 55 on its Intelligence Index (version 5.0.1), well above the 37 median for models in its price tier. The pattern in Anthropic's own table is consistent: Sonnet 5.5 is dramatically better than Sonnet 5 everywhere, and close to Opus 5.5 on most tests — with one headline exception in its favour.

*Opus 5.5's Terminal-Bench figure is at Xhigh effort. Sonnet 5.5 ran at Max effort unless a row says otherwise.

The Terminal-Bench jump is the story of the release. Sonnet 5 scored 10.3% on the same test; Sonnet 5.5 scores 70.6%, above Opus 5.5's 66.4%. On knowledge work the gap to the flagship nearly vanishes: GDPval-AA puts Sonnet 5.5 at 1844 against Opus 5.5's 1846, about 400 points above Sonnet 5. On computer use it climbs from 57.0% to 80.1% on OSWorld 2.1, within two points of Opus. Anthropic says it is also the first Sonnet model to beat Pokémon Red working only from screenshots, a small detail that says more about its visual grounding than any chart does.

The honest reading is that Sonnet 5.5 is not a smaller Opus. Anthropic says so itself: "Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment." On FrontierCode it trails Opus by eight points at Max effort, and on Humanity's Last Exam by three. What changed is the floor: work that used to need the flagship now fits comfortably in the mid-tier model, at half the per-token price.

What It Costs, and the Effort-Level Catch

The headline price is unchanged from Sonnet 5: $2 per million input tokens and $10 per million output tokens. Converted at November 8, 2026 mid-market rates of 1 USD = 0.7471 GBP and 1 USD = 1.41 CAD, that is about £1.49 / C$2.82 per million input tokens and £7.47 / C$14.10 per million output tokens. Sterling and Canadian figures are approximations. The price list around it is where the money actually goes:

Two things in that table need a note. First, the cache-read price: Anthropic's pricing table still shows $0.20, but the company announced a cut to $0.10 on October 7, 2026, alongside the Haiku 5.5 launch, and its own pricing page says cache reads on Sonnet 5.5 are charged at 0.05x the input rate, which is $0.10. Budget on $0.10; the table cell looks stale. Second, if you route inference through the US-only inference_geo option, a 1.1x multiplier applies to every token price above.

The bigger catch is the effort setting. Sonnet 5.5 lets you pick how hard it thinks: Low, Medium, High, Xhigh, and Max. Claude Code and the Claude apps default to Medium; the Platform defaults to High. At Low or Medium, Anthropic's own charts show Sonnet 5.5 beating Sonnet 5's best scores for about a tenth of the cost per task. At Max, the maths flips. Artificial Analysis data, as analysed by eesel, puts Sonnet 5.5 at $7.60 per task at Max effort against $5.98 for Opus 5.5 at Max, because Sonnet burns about 193,000 tokens per task where Opus uses 119,000. Anthropic's own post concedes the shape of this: Sonnet 5.5 "complements Opus 5.5 best when running at lower effort settings," and "at higher settings, it can perform comparably at a similar cost."

It can also score worse at Max. On FrontierCode, Sonnet 5.5 scores 52.1% at Xhigh but only 46.2% at Max, because at Max it more often ran Claude Code's code-review skill, which led to timeouts or edits beyond the task's scope. More thinking made the result worse. One independent tester, Simon Willison, ran the same SVG prompt at every effort level on launch day: Low cost 1.6 cents and took 10 seconds; Max cost $1.28, ran for 15 minutes 40 seconds, burned its full 128,000 thinking tokens, and failed to produce the image. The rule that follows is simple: treat Max as off-limits for Sonnet 5.5, and hand anything that needs that much reasoning to Opus 5.5.

None of this changes the fact that the model is cheaper to run than its predecessor at the same list price, because it uses fewer tokens to do the same work. Balyasny Asset Management ran 2,441 finance tasks and found Sonnet 5.5 used about 121,000 tokens per answer where Sonnet 5 used 497,000. Slack reported about 14% fewer output tokens with no prompt changes, Lovable a third fewer tool calls, and Zendesk said tickets were processed 20% faster with fewer wrong escalation decisions. Those are Anthropic-reported customer results, not independent tests, but they all point the same way.

For chat rather than the API, the plans are the same ones that cover the rest of Claude. Free costs $0 with limited usage. Pro is $20 a month, or $17 a month when billed annually ($200 up front). Max is $100 a month for 5x Pro usage plus $100 a month in API credits. Ultra is $200 a month for 10x weekly usage, no 5-hour limit for verified accounts, and $200 a month in API credits. Enterprise is $20 per seat per month plus usage billed at API rates. All prices exclude tax. If you only want to try the model, the Free plan costs nothing and the API bills pay-as-you-go; there is no separate charge for Sonnet 5.5 itself.

The 1M Context Window

Sonnet 5.5's context window is 1 million tokens, and Anthropic charges the full window at standard pricing — a 900,000-token request is billed at the same per-token rate as a 9,000-token request, with no long-context surcharge. That is the same window Opus 5.5 and Haiku 5.5 carry, and it is enough to hold an entire codebase, a lengthy contract, or dozens of research papers in a single request.

What matters more than the size is whether the model reasons well across all of it. Anthropic points to its Vending-Bench Arena result, where Sonnet 5.5 ran a simulated business over time and developed a strategy of investing heavily in capacity for the first ten simulated months before pivoting sharply to profitability — and finishing well ahead of the competition. On the practical side, the 512-token minimum cacheable prompt means shorter system prompts now qualify for cache reads, and the Batch API supports up to 300,000 output tokens with the output-300k-2026-03-24 beta header. The Batch API itself carries a 50% discount, at $1/$5 per million tokens.

Safety: The First Sonnet With Opus-Level Cyber Safeguards

Sonnet 5.5 is the first Sonnet model to launch with cybersecurity safeguards like those on Opus 5.5. Anthropic says its cyber capabilities are comparable to Opus 5's, so the safeguards follow. Routine software development, including finding and fixing bugs in your own code, is unaffected, but higher-risk cybersecurity tasks will visibly fall back to Sonnet 5. Security teams doing authorised defensive work can apply to Anthropic's expanded Cyber Verification Program for tiered access to more advanced capabilities on Sonnet 5.5, Opus 5.5, and the Mythos models.

The biology safeguards are the same as Sonnet 5's: they target a narrow set of high-risk requests, and most research, education, and clinical work is unaffected, though Anthropic admits some microbiology and virology requests may be flagged in error. A Life Sciences Verification Program exists for organisations that need the full breadth of biology work.

Two newer protections arrive with this model. Sonnet 5.5 is the first Sonnet to launch with safety classifiers that prevent reasoning extraction, the distillation attacks where thousands of fake accounts are used to pull a model's capabilities out at industrial scale, and it expands preserved thinking, so Claude's thinking cannot be decoupled from the account that created it. On Anthropic's automated behavioural audit, covering roughly 1,850 scenarios, Sonnet 5.5 improves on or matches Sonnet 5 on most measures of alignment, resistance to misuse, and honesty, and it comes close to Opus 5.5 on containment tests measuring how rarely it tries to escape its sandbox. Anthropic also notes, candidly, that no set of evaluations reliably catches every failure — so these results are a floor, not a guarantee.

Sonnet 5.5 vs Opus 5.5 vs Haiku 5.5: Which One to Pick

The three models in the 5.5 family are priced 20x apart from top to bottom of the mid-range, and Anthropic's own guidance is blunt about which job each one fits. Haiku 5.5 launched on October 7, 2026, and our Claude Haiku 5.5 review covers it in full.

*Haiku 5.5's $0.10/$0.50 rate applies to prompts up to 100,000 tokens; longer prompts cost $0.50/$2.50. Anthropic says around 90% of requests to the previous Haiku model fell under the threshold.

The practical split is the one Anthropic itself draws: Sonnet 5.5 and Opus 5.5 remain the better choices for complex agentic coding, while Haiku 5.5 fits narrowly scoped tasks that would otherwise be cost-prohibitive, like compaction, summarisation, or subagent work. In Claude Code, Sonnet 5.5 is the default model at Medium effort, which is where most developers will meet it first. Our Claude Code explainer covers how the tool itself is priced and where its usage limits sit.

Common Mistakes to Avoid

Do not run Sonnet 5.5 at Max effort and expect Opus results at Sonnet prices. The independent data says the opposite: at Max it costs more per task than Opus 5.5 and can score lower, because it burns far more tokens. Low, Medium, or High is where the value sits. Do not quote the $2/$10 price as the whole cost without the cache and effort context. Cache writes at $2.50 or $4 per million tokens, and the effort level you pick, move the bill more than the base rate does. Do not assume the $0.20 cache-read figure on Anthropic's pricing table is current; the company cut it to $0.10 on October 7 and the table has not caught up. Do not migrate code from Sonnet 5 without reading the migration guide. thinking: disabled, forced tool_choice, and non-default sampling parameters all return 400 errors now, and the old computer-use tool version has been replaced. And do not treat the benchmark table as an independent verdict. Every score in it is Anthropic-reported, and the one outside check, Artificial Analysis, scores Sonnet 5.5 at 55 on its index rather than near the top of its board.

FAQ

How much does Claude Sonnet 5.5 cost?

$2 per million input tokens and $10 per million output tokens on the API, the same as Sonnet 5. Cache reads cost $0.10 per million tokens, cache writes $2.50 (5-minute) or $4 (1-hour), and the Batch API halves both base rates to $1/$5. In chat, it is included in the Free, Pro ($20/month), Max ($100/month), and Ultra ($200/month) plans.

Is Sonnet 5.5 better than Opus 5.5?

On some tests, yes: it beats Opus 5.5 on Terminal-Bench 4.0, 70.6% against 66.4%, and matches it on GDPval-AA, 1844 against 1846. On complex, open-ended work requiring sustained judgment, Anthropic says Opus 5.5 remains clearly stronger, and at Max effort Opus costs less per task. Pick Sonnet for well-scoped work and Opus for the hardest reasoning.

What is the 1M context window?

Sonnet 5.5 accepts up to 1 million tokens of input in a single request — enough for an entire codebase or dozens of research papers — at standard per-token pricing with no long-context surcharge. Maximum output is 128,000 tokens, or 300,000 on the Batch API with a beta header.

Can I use Sonnet 5.5 for free?

Yes, through the Free plan on claude.ai, which includes limited usage every 5 hours and every week. The API is pay-as-you-go with no free tier, though Max and Ultra subscribers get $100 and $200 a month in API credits respectively, usable on any model.

Does Sonnet 5.5 replace Sonnet 5?

Effectively, yes. It is faster, cheaper per task, and stronger on every benchmark Anthropic published, at the same $2/$10 list price. Sonnet 5 remains available, but there is no pricing or capability reason to choose it for new work.

When was Claude Sonnet 5.5 released?

September 28, 2026, six days after Opus 5.5. Haiku 5.5, the third model in the family, followed on October 7, 2026.

Sonnet 5.5 is the first mid-tier Claude model that makes the flagship feel optional for most real work. It beats Opus 5.5 on the coding benchmark that matters most to developers, matches it on knowledge work, and costs half as much per token, provided you keep it at Low, Medium, or High effort and let Opus handle the jobs that need Max. If you build with Claude, switch the well-scoped work to Sonnet 5.5 now and keep Opus 5.5 for the open-ended judgment calls. That split is where the value is.

Sources

• Anthropic — Introducing Claude Sonnet 5.5
• Anthropic — Introducing Claude Haiku 5.5
• Claude Platform Docs — Claude Sonnet 5.5 overview
• Claude Platform Docs — Pricing
• Claude — Plans and pricing
• Artificial Analysis — Claude Sonnet 5.5
• eesel — Claude Sonnet 5.5 review: benchmarks, real costs, and the verdict
• MarkTechPost — Anthropic Releases Claude Sonnet 5.5
• Atomic Agent — Sonnet 5.5 vs Opus 5.5: Price, Benchmarks, Effort Levels
• OmniaKey — Claude Sonnet 5.5 Pricing and Cost