Anthropic Releases Claude Sonnet 5.5, the Second Model in the Claude 5.5 Series Following Opus 5.5
Anthropic has launched Claude Sonnet 5.5, the second model in the Claude 5.5 series following Opus 5.5. The company claims that the output is over 30% faster than Sonnet 5, and it can save up to about 30% on similar tasks. The four tiers are Fable, Opus, Sonnet, and Haiku, with prices decreasing as you go down: Fable targets large projects that can run for hours across codebases, Opus 5.5 is for complex tasks that require repeated trade-offs, Sonnet 5.5 is aimed at well-defined daily tasks like bug fixing, documentation, and presentations, while Haiku 5.5 is expected to be released in a few weeks.
Programming shows the largest increase. On Terminal-Bench 4.0, Sonnet 5.5 scores 70.6%, Sonnet 5 scores 10.3%, and Opus 5.5's highest tier scores 66.4%. On CursorBench 4.0, it scores 55.5%, compared to Sonnet 5's 34.1% and Opus 5.5's 57.8%. FrontierCode 1.1 (Max) scores 46.2%. On the knowledge work benchmark GDPval-AA v2.1, it scores 1844, while Opus 5.5 scores 1846 and Sonnet 5 scores 1449. Internal tests assigned a 10-page business review to the model using quarterly reports, call records, and presentation templates, and two experts deemed the draft ready for external release. The company claims it is the first Sonnet to complete Pokémon Red solely through screenshots. The thinking intensity ranges from Low to Max in five tiers, with the application and Claude Code defaulting to Medium; the company cites that on Terminal-Bench, Sonnet 5.5 uses Medium at about $0.83 per task, scoring 28.8%, while Sonnet 5 at Max costs about $11.62 per task, scoring only 10.3%. The company reminds that for open-ended tasks requiring continuous judgment, Opus 5.5 remains significantly stronger. Independent evaluation agency Artificial Analysis provided Terminal-Bench scores that do not completely align with the official figures.
The pricing is the same as Sonnet 5: $2 per million input tokens, $10 for output, and $0.20 for cache reads, which is half of Opus 5.5; savings come from using fewer tokens for the same work and more frequently merging multi-step tool calls. In terms of security, its cybersecurity capabilities are rated on par with Opus 5, making it the first Sonnet with cybersecurity protection: routine bug checks are unaffected, while high-risk requests are redirected to Sonnet 5 with a prompt to switch. It also becomes the first Sonnet with an anti-distillation mechanism, binding the thinking content to the generating account, meaning that switching accounts mid-way in Claude Code will affect performance. The model is now available on Anthropic products and on AWS, Google Cloud, and Microsoft Azure, with the API named claude-sonnet-5-5; old calls that disable the thinking function need to switch to between_tools. Sellers are developers treating mid-tier models as the default production entry, while buyers are enterprises aiming to match flagship daily scores at half the price. The event is driven by the completion of mid-tier models in the 5.5 series, with funding shifting from high-priced Opus calls to the faster Sonnet default tier. Beneficiaries are programming and office workflows billed by task, while those under pressure are more expensive models that still rely on flagship premiums to cover the same batch of daily tasks.
Source: Public Information
ABAB AI Insight
Anthropic has delegated programming agent capabilities from the flagship to mid-tier pricing, changing the default workflow rather than just issuing another performance poster. The Terminal-Bench scores from 10.3% to 70.6%, intersecting with Opus 5.5's 66.4%, indicating that "daily agent coding" no longer needs to be at the highest price tier. CursorBench is only two points apart, corresponding to real editor sessions, making it closer to procurement scenarios than synthetic tasks. The GDPval-AA score of 1844 compared to 1846 nearly levels knowledge work, with Fable still reserved for ultra-long projects. The four-tier structure has become: Fable buys time span, Opus buys judgment, Sonnet buys throughput, and Haiku buys volume.
The capital path is the same $2/$10 pricing, fewer tokens, and faster output, shifting gross profit from "expensive models with fewer calls" to "mid-tier models with high concurrency." Merging tool calls reduces steps, directly impacting bills settled by tokens. The anti-distillation mechanism binds thinking trajectories to accounts, countering the reasoning chains of bulk fake accounts; high-risk cybersecurity requests are redirected back to Sonnet 5, removing the attack surface from the new weights. The simultaneous launch on AWS, Google Cloud, and Azure sends mid-tier models into existing procurement contracts, rather than discussing another round of model replacements.
This is analogous to how Sonnet 3.5 captured a large share of programming at mid-tier prices, and how OpenAI used mini/flash to divert default chat traffic. The industry phase has shifted from "who scores highest" to "who becomes the default engine for IDEs and office suites." Independent evaluations do not completely overlap with official rankings, indicating that effort tiers, fallback strategies, and tool cycles will affect rankings; procurement will still consider dollars per task rather than just peak percentage scores.
Structural judgment belongs to the technical replacement combined with the transfer of pricing power. The mechanism is that once the capability curve for agent coding is leveled at mid-tier, flagship premiums can only cover open-ended long-term tasks. Whoever controls the default tier controls the token flow. Security classifiers and anti-distillation have transformed mid-tier models from interfaces that can be freely distilled into controlled runtimes bound to accounts. If Haiku 5.5 further cuts prices, the default tier will continue to shift downwards.
ABAB News · Cognitive Laws
Mid-tier catching up to flagship daily scores will lead to a change in default traffic tiers
Stable pricing with fewer tokens is the true price reduction
Anti-distillation binding to accounts recovers reasoning chains from public goods