Claude Sonnet 5.5 Is Second on the Index at Half the Price. At Max Effort It Spends the Difference.

Anthropic shipped Claude Sonnet 5.5 today, September 28, 2026, six days after Claude Opus 5.5 opened the 5.5 family. The pitch is the one every mid-tier launch makes: most of the flagship's ability at a fraction of the price. Anthropic's own line is "a clear upgrade over Claude Sonnet 5, runs 30%+ faster, and costs up to 30% less for most work."

Two of those three claims hold on independent data. The third holds only if you stay off the top effort setting. Artificial Analysis scored the model this afternoon at 56 on its Intelligence Index, second of 217 and two points behind Opus 5.5, and then measured something no other model has produced: about 193,000 output tokens per index task at max effort. Half the per-token price, roughly double the tokens. We read the announcement, the pricing page, the migration guide, the prompting guide, the system card, and AA's analysis. Here is what held up.

Same Price as Sonnet 5, Smarter, and Not Cheaper at the Top

The price did not move. The pricing page row is $2 and $10, the same as Sonnet 5, with a 1M-token context at that rate. Anthropic's Sonnet 5 "introductory" price never lapsed; the $3/$15 increase once scheduled for September 1 did not happen, and Sonnet 5 is now listed as active-legacy at the same $2/$10 with retirement not sooner than June 30, 2027.

The cutoff moved five months. June 2026 against Sonnet 5's January 2026. For anyone who asks a model about this year's API changes, that is the quiet win in the table.

Thinking cannot be turned off. Sending thinking: {"type": "disabled"} returns a 400. The replacement is a new between_tools setting that suppresses up-front reasoning but still emits progress blocks between tool calls, and it is accepted only at low, medium, and high effort. Ask for it at xhigh or max and you get another 400.

Five breaking changes versus Sonnet 5, per the what's-new page: the disabled-thinking removal above; forced tool use (tool_choice of any or a named tool) now returns 400, so use auto plus structured outputs; thinking blocks are bound to the model, the conversation prefix, and the account, with history edits before a block rejected by default for accounts created on or after August 31, 2026; the computer_20251124 tool is rejected on the Claude API and Google Cloud; and temperature, top_p, top_k, and assistant prefill are gone. There is also a response-shape change that is not on the breaking list: text the model emits between tool calls now arrives inside thinking blocks, which are empty at the default display setting. If your agent loop reads progress text from content blocks, it will read nothing.

The hidden line item. The prompting guide says it plainly: "Thinking counts toward max_tokens even when thinking content isn't returned to you," and "From medium up, the model thinks briefly before almost every reply, even a greeting." The API default is high. A plain caller who never sets effort pays for reasoning tokens on every call and never sees them. Changing the top-level effort between requests also invalidates the prompt cache.

The Benchmarks, Sorted by Who Reported Them

House rules: independent numbers first, stated as fact with the source and the index version; Anthropic's numbers labeled as Anthropic's, with the effort level, because the vendor table mixes them.

Artificial Analysis, Intelligence Index v4.3.2, launch-day analysis. Sonnet 5.5 (max, with Anthropic's default server-side fallback) scores 56, second of 217, behind Opus 5.5 (max) at 58. The effort ladder runs 52 at xhigh, 47 at high, 41 at medium, 36 at low. Sonnet 5 (max) reads 38 on the same scale, so even Sonnet 5.5 at medium outscores Sonnet 5 at max. On AA's own Terminal-Bench 4.0 run, Sonnet 5.5 posts 64% against 60% for both Opus 5.5 and GPT-6 Astra. On the two real-work Elo boards it sits just under the flagship: GDPval-AA v2.1 at 1844 against Opus 5.5's 1846, AA-Briefcase v1.1 at 1811 against 1822.

The cost row is the story. AA's article puts the token figure at roughly 60% above both Opus 5.5 (max) and Sonnet 5 (max), and about seven times GPT-6 Astra (max). At $10 per million output tokens, 193,000 tokens is $1.93 of output on one task before any input. That is how a model priced at half of Opus 5.5's $4/$20 ends up costing about 50% more per task than its own predecessor. The announcement's "up to 30% less per task" is measured, Anthropic says, "in our testing"; it is not what AA measured at max.

Now the vendor table. These are Anthropic's self-reported numbers from the announcement, run with adaptive thinking at max effort and averaged over five trials unless footnoted. We show Sonnet 5.5, its predecessor, and Opus 5.5.

Four honest observations the announcement does not lead with:

• The Terminal-Bench win is inside the noise. 70.6% against Opus 5.5's 66.4% is the headline row, and the system card is careful about it: the run was Claude Code in bare mode at max effort with safeguards on, 1.2% of requests were answered by the fallback model, and the standard error is about 2.5 points. The gap is under two standard errors. AA's own harness agrees on the direction (64 vs 60). We print "edges Opus 5.5, within noise" and nothing stronger.
• Max is not the best setting on two rows. FrontierCode scores 52.1% at xhigh and 46.2% at max; Anthropic's footnote says the max run launched Claude Code's code-review subagents and lost points doing it. GDPval-AA drops from 1844 at max to 1725 at xhigh, but with about 67% fewer output tokens. The effort knob is a cost knob and sometimes a quality knob, and not always in the same direction.
• It trails Opus 5.5 on every coding row except the terminal. SWE-Bench Pro is 81.3% against 89.9%, FrontierCode 46.2% against 54.4%, CursorBench 55.5% against 57.8%. The system card's own summary says the model is "broadly less capable than Opus 5.5 across domains."
• The speed claim is conservative. Anthropic says "30%+ faster than Sonnet 5." AA's launch measurement has Sonnet 5.5 generating output at about 139 tokens per second at max against 79 for Sonnet 5, which is closer to 75% faster. Time to first answer token at max is another matter: AA clocks it at 402 seconds, because that is reasoning time, and 1.26 seconds at low.

Where the Safeguards Reroute You

Sonnet 5.5 carries the three-stage cyber safeguards from Opus 5.5 and is the first Sonnet with reasoning-extraction classifiers, Anthropic's anti-distillation filter. The fallback behavior differs by category, and the system card spells it out: cyber and frontier-LLM blocks fall back to Sonnet 5 and answer anyway; chemical and biological, reasoning-extraction, and conventional-weapons blocks "have no fallback model." On the consumer apps, a help-center article says a tripped request "re-runs on Sonnet 5 in the same conversation" with a visible notice.

The card's other numbers worth a buyer's attention: no new Responsible Scaling Policy thresholds; alignment risk assessed as low; the lowest sandbox-escape rate of any recent Claude at 9 of 196 runs; and a large drop in prompt-injection susceptibility on Gray Swan's Shade coding attacks, 3.01% attack success against 19.47% for Sonnet 5 with thinking and 54.61% for Opus 5.5 without probes. The card also calls Sonnet 5.5's thinking the least legible of the family, which matters if you audit reasoning traces.

Where You Can Run It

Claude API for all customers as claude-sonnet-5-5; Amazon Bedrock as anthropic.claude-sonnet-5-5, access-gated behind "See Access" and limited to the Standard service tier with no structured outputs; Claude Platform on AWS; Google Cloud's Agent Platform on global and multi-region endpoints only, with a 10% premium on multi-region and regional and a 30 MB payload cap; Microsoft Foundry as Global Standard only, billed in Claude Consumption Units. Oracle's catalog has no Anthropic model. Zero data retention is available. GitHub Copilot added it today for Pro, Pro+, Max, Business, and Enterprise with a gradual rollout. Press reports put it as the new Claude Code default on the Anthropic API from today, with the sonnet alias still resolving to older models on Bedrock, Google Cloud, and Foundry; we have not read the Claude Code changelog ourselves, so treat the version detail as press-sourced.

Consumer-side, this is the model on the Free plan. Anthropic's product page says anyone can chat with Sonnet 5.5 on claude.ai, web, iOS, and Android; Pro, Max, Team, and Enterprise get it too, at medium effort by default. Opus 5.5 stays off Free.

Verified vs Unconfirmed: The Scorecard

Who Should Use It

Use Sonnet 5.5 if you run Sonnet 5 today and can set effort yourself. At high, medium, or low it outscores Sonnet 5 at max on AA's index while spending a fraction of the tokens, it is markedly faster, the cutoff is five months fresher, and it is the most prompt-injection-resistant Claude on the card. For coding agents with a fixed budget, set effort to high or xhigh, set max_tokens to 128,000 with room for thinking, and stream.

Skip it for anything that assumes thinking can be switched off, for pipelines that read inter-tool progress text from content blocks, and for max-effort runs where the bill matters: at max, AA's per-task cost is about 50% above Sonnet 5 and close to Opus 5.5, which beats it on nearly every coding row. If a request can trip the cyber classifier, remember that the answer may come from Sonnet 5.

The frontier is elsewhere: Claude Opus 5.5 (max) still leads the v4.3.2 index at 58, with Claude Fable 5.1 and GPT-6 Astra at 53. Sonnet 5.5 at 56 is the best mid-tier model anyone has shipped, and the flagship is two points and $2 per million output tokens away. Our model comparison has the current standings side by side.

Update, October 16, 2026

• October 7: Anthropic cut Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens, "0.05x the base input price instead of 0.1x", in a release-notes entry published inside the Claude Haiku 5.5 launch. Every other price is unchanged. Anthropic also edited the September 28 announcement to show $0.10 with no correction note; a Wayback Machine capture from September 30 still reads $0.20, as did AA, SiliconANGLE, and The Decoder on launch day. Anthropic's estimate that the cut makes most agentic work "around 20% cheaper" is a vendor estimate with no workload stated; under AA's 7:2:1 cache-to-input-to-output blend the saving is 4.5%.
• September 29: GPT-6.1 Sol replaced GPT-6 Sol, so the GPT-6 Sol column in Anthropic's table was stale within a day. AA scores GPT-6.1 Sol at 52.
• September 30: Anthropic deprecated Claude Sonnet 4.5, naming Sonnet 5.5 as the replacement, with retirement on November 30, 2026.
• October 1: Epoch AI's Capabilities Index put Opus 5.5 at 167 and said Sonnet 5.5 "roughly matched Claude Fable 5.1 (165)."
• October 8: The New Stack ran three coding tasks five times each at max effort: Sonnet 5.5 passed 15 of 15 runs to Opus 5.5's 13 of 15, at $12.69 against $22.07.
• October 16: AA's live model page now prints $5.46 per index task at max (from $7.60 on launch day) and $0.48 at medium (from $0.59), with no changelog note, while Opus 5.5 ($5.98) and Sonnet 5 ($5.09) did not move; the cache cut explains only about 4.5% of it. The verbosity ranking stands: 420M output tokens to run the index, last of 115 models. Vals AI ranks Sonnet 5.5 second of 48 at 67.04%, a hair above Opus 5.5's 66.97%; Zapier's AutomationBench 1.0.6 has it third at 44.75%, ahead of Opus 5.5 and behind Gemini 4 Argon. On AA's GDPval-AA board the pair now reads 1838 against 1866, with Meta's Muse Wonder Full between them at 1857.

Sources

• Anthropic: Introducing Claude Sonnet 5.5
• Anthropic pricing page
• Anthropic docs: What's new in Claude Sonnet 5.5
• Anthropic docs: Prompting Claude Sonnet 5.5
• Claude Sonnet 5.5 System Card (PDF)
• Anthropic developer platform release notes
• Artificial Analysis: Claude Sonnet 5.5 launch analysis
• Artificial Analysis: Claude Sonnet 5.5 model page
• Vals AI: Claude Sonnet 5.5

*Not sure whether a mid-tier model at the right effort setting covers your work, or whether you need the flagship? Take the free 2-minute quiz and get matched. Then read what Anthropic shipped nine days later at a tenth of the price, with a surcharge of its own: Claude Haiku 5.5.*