Claude Sonnet 5.5

Anthropic is shipping Claude Sonnet 5.5 six days after Opus 5.5 and three months after Sonnet 5. Same $2 / $10 price as Sonnet 5, 30%+ faster output, and a benchmark table where the cheap model beats Opus 5.5 on the agentic coding test.

What shipped

Model type: Text + image input, text output

Pricing: $2 / $10 per million input / output tokens, identical to Sonnet 5 and half of Opus 5.5’s $4 / $20. Cache reads are $0.20 per million (the same as Opus 5.5, which makes them 10% of input here against 5% on Opus), cache writes $2.50 (five-minute) and $4 (one-hour), Batch $1 / $5.

Available on: claude.ai and the Claude apps, Claude Code, Claude Cowork, the Claude Developer Platform, Amazon Bedrock and Claude Platform on AWS, Google Cloud, and Microsoft Foundry. The less-restricted cyber build comes “in the near future” through the expanded Cyber Verification Program.

Other info: 1M-token context window, 128K max output. Anthropic says output generates “30%+ faster than Sonnet 5, making it our fastest Sonnet model to date.” Reliable knowledge cutoff and training data cutoff both June 2026. Same tokenizer as Sonnet 5. Minimum cacheable prompt drops to 512 tokens from 1,024. It’s the first Sonnet to beat Pokémon Red working only from screenshots. The 148-page system card calls itself “condensed” and says future non-frontier cards will be too.

More details: Introducing Claude Sonnet 5.5, the system card, and What’s new in Claude Sonnet 5.5 for developers.

Anthropic released Claude Sonnet 5.5 on Monday as “the second model in the Claude 5.5 family,” pitched as “a clear upgrade over Claude Sonnet 5” that “runs 30%+ faster, and costs up to 30% less for most work.” The positioning is a division of labor. Opus 5.5 is “built for complex work requiring careful judgment,” and Sonnet 5.5 is “strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets,” with “a sharp eye for design” thrown in. It costs the same per token as Sonnet 5, runs on every Anthropic surface and all four clouds on day one, and carries over most of the Opus 5.5 API contract: no forced tool use, thinking blocks bound to the conversation and the account, progress notes between tool calls moved into thinking blocks, and the new computer use toolset required on the Claude API and Google Cloud.

The evidence is a bit lopsided. Sonnet 5.5 beats Opus 5.5 on Terminal-Bench 4.0, Terminal-Bench-Science, AutomationBench, HealthBench Professional (69.2% vs 65.6%), and BenchCAD, and lands within two Elo points on GDPval-AA. Anthropic’s answer is that “benchmark scores capture only one facet of a model’s capabilities” and that “Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment.” The cost charts make the stronger case: at Medium, the default in the Claude apps, it “far exceeds Sonnet 5’s best score for less than a tenth of the cost per task” on Terminal-Bench, and at High on FrontierCode it scores 10 points over Sonnet 5 “at about one fifteenth of the cost per task.” Partner numbers follow the same line. Balyasny says it used about 121k tokens per answer where Sonnet 5 used 497k across 2,441 finance tasks, Slack measured about 14% fewer output tokens, Base44 reports 3.6 iterations per app build against Opus 5’s 7.7, and Zendesk says tickets moved 20% faster. The catches are in the system card. The model “hallucinates more” than Opus 5.5, its FrontierCode score drops at max effort, and on Toolathlon it averaged 31.6 turns per task, more than any Claude model in the table (which sits awkwardly next to all those fewer-steps quotes).

What’s new

Most of the release is Opus 5.5’s machinery pushed down to the $2 tier. Four changes matter if you run Sonnet in production.

• Opus-grade agent scores at Sonnet prices. Sonnet 5 posted 10.3% on Terminal-Bench 4.0 and 10.7% on AutomationBench. Sonnet 5.5 posts 70.6% and 44.7%, beating Opus 5.5 on both at half the per-token price. If your Opus bill is mostly well-specified agent work, you’re overpaying as of this morning.

• You can still turn thinking off, sort of. Opus 5.5 killed the off switch. Sonnet 5.5 keeps one as thinking: {"type": "between_tools"}, which skips up-front thinking at low, medium, or high effort while still writing short progress notes between tool calls. Send the old disabled and you get a 400 that points you to the new value. It’s the only way left in the 5.5 family to skip the up-front think.

• The first Sonnet that needs cyber classifiers. Sonnet 5 never built a working Firefox exploit in Anthropic’s June eval. Sonnet 5.5 hit full arbitrary code execution in 43.4% of ExploitBench runs, so it ships with the same three-stage cyber gate as Opus 5.5 and visibly hands flagged requests to Sonnet 5 in Anthropic’s apps. On the API the fallback is opt-in, and without it a flagged request just returns a refusal.

• Cheaper to cache small prompts, easier to steer long runs. The 512-token cache minimum (down from 1,024), per-message effort, mid-conversation system messages, and mid-conversation tool changes all arrive on Sonnet for the first time. None of that shows up on a leaderboard. All of it shows up on the invoice of anything that runs long agent loops.

The benchmarks

Anthropic’s numbers

Max effort, five-trial averages unless noted.

• Terminal-Bench 4.0: 70.6%. +4.2 over Opus 5.5 (66.4% at xhigh), +12.7 over GPT-6 Astra (57.9%), +60.3 over Sonnet 5 (10.3%).

• AutomationBench (Zapier): 44.7%. +2.2 over Opus 5.5 (42.5%), +12.7 over GPT-6 Sol (32.0%), +34.0 over Sonnet 5 (10.7%).

• GDPval-AA v2.1: 1844 Elo. −2 under Opus 5.5 (1846), +109 over Fable 5.1 (1735), +357 over GPT-6 Sol (1487), +395 over Sonnet 5 (1449).

• FrontierCode v1.1: 46.2% (52.1% at xhigh). −8.2 under Opus 5.5 (54.4%), −3.1 under GPT-6 Sol (49.3%), +3.8 over Sonnet 5 (42.4%). Max effort kept running Claude Code’s review skill and wandered out of scope.

• OSWorld 2.1: 80.1% partial. −1.7 under Opus 5.5 (81.8%), +23.1 over Sonnet 5 (57.0%).

How and where to use it

Where it’s available

• Claude.ai and the Claude apps

• Claude Code, and Claude Cowork

• The Claude Developer Platform plus Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry

What it’s good at

• Terminal-heavy agent work and business automation

• Bug fixing and code review in well-scoped repos (CodeRabbit is moving “simple and moderate reviews” over now)

• Support and ticket workflows, per Zendesk’s “fewer wrong decisions” and 20% faster resolution

• Finance and document extraction at volume, where Balyasny called it “the best quality-to-cost tradeoff of the seven models we ran”

• Decks and docs from a template

• Charts and CAD with tools (90.2% on Chartography, 0.963 on BenchCAD). Healthcare Q&A

• Anything latency-sensitive

What it’s bad at / shouldn’t be used for

• Long-context reconstruction work

• Open-ended, judgment-heavy work

• Anything where a confident wrong answer costs you, since it hallucinates more than Opus 5.5

• Penetration testing, exploit development, and binary vulnerability scanning.

• Harnesses that force tool calls with tool_choice, edit history mid-session, or pair Sonnet with an Opus 4.8 or Sonnet 5 advisor

• GUI computer use against untrusted pages

First impressions

The announcement is barely fifteen minutes old, so this is the first wave from the Hacker News thread, already past 30 points.

The positives

The first comment did the reading for everyone. User ramish94:

“Opus 5.5 might be the best model I’ve ever used and Sonnet 5.5 matches it and exceeds in some benchmarks. Clearly Anthropic have had some sort of breakthrough with not just performance but also cost with the 5.5 family”

The “breakthrough” framing is doing more work than the evidence, since Anthropic’s own page spends a paragraph telling you Opus is still the better model where it counts.

The most useful reaction was a backhanded one. User bigyabai replied:

“It’s long overdue. Sonnet 5 was terrible API value for agentic coding, there were open models like GLM-5.3 Flash that blew it out of the water at 1/20th of the price.”

10.3% on Terminal-Bench 4.0 for the model it shipped as its most agentic Sonnet. The positive case for 5.5 is partly that it ends the stretch where the cheap Claude was the wrong answer on price.

User SubiculumCode added the product angle:

“Yeah, I did kind of feel like the step down from Opus 5.5 was so large as to never make it appealing.”

The negatives

“Paying $200 a month and part of their Cyber Verification Program but can’t use Opus 5.5 or Sonnet 5.5 for any authorized bounty work. Immediately get flagged or Cyber.”

User AshamedBadger56 replied that the program “does absolutely nothing.” That lines up with the system card, which says the less-restricted Sonnet 5.5 build comes “in the near future” and warns about more refusals on benign security work. The Fable 5.1 and Opus 5.5 launches promised the same CVP timing.

User pookieinc spotted who’s missing from the chart:

“It’s interesting that in all their benchmarks, they omit Fable numbers and only focus on Opus, Sonnet, and OpenAI models. Maybe Fable is out the door?”

The launch table compares against Sonnet 5, Opus 5.5, and GPT-6 Sol only. Fable 5.1 shows up in the system card (Sonnet 5.5 beats it on Terminal-Bench, GDPval, and AA-Briefcase), so the omission reads as positioning. Anthropic is selling Sonnet against Opus now, and Fable is the model nobody’s pitching this week.

And the question every launch like this raises. User protoman3000:

“Since Opus 5.5 I’ve been barely using Sonnet anymore. It interestingly enough doesn’t consume that many tokens.”

UncleOxidant agreed that their Pro quota stopped running dry after Opus 5.5 shipped. If Opus 5.5’s token efficiency already made it cheap enough for a subscriber to use all day, Sonnet 5.5 has to win on speed and API cost rather than on quota, and that’s a narrower pitch than the benchmark table suggests.

Jake’s take

Sonnet has been a frustrating model line for me. Every iteration is “ok” on most tasks, and I never just want “ok”; especially when Opus is sitting right there. There are rarely non-demanding tasks I want these models to do; I’ll just do those myself. Now that Anthropic has a four model line, I suspect this issue will get worse and worse. Anything brainless, I’ll send to Haiku. Anything that needs some juice, I’ll send to Opus. Anything above that, Fable. So what do I need Sonnet for anymore?

Also — the benchmark table is a bit suspicious. Three months ago Anthropic sold Sonnet 5 as “the most agentic Sonnet model yet,” and today’s table has that model at 10.3% on Terminal-Bench 4.0 and 10.7% on AutomationBench. Sonnet 5.5 supposedly beats Opus 5.5 on this metric (we’ll see), but the same page tells us to ignore that because Opus is “clearly stronger” at the work that matters.

If the flagship and the mid-tier can’t be separated by the published numbers, the published numbers stopped doing their job. The “up to 30% less” claim rests entirely on token efficiency at an unchanged list price, and the system card’s own Toolathlon run has it taking more turns than any Claude model in the table while the alignment section notes it produces “more output tokens per transcript” than earlier models. Also, “hallucinates more” in a system card is never something I want to do.

Sorry Anthropic, Opus 5.5 is too good to care about this one.