Anthropic released Claude Sonnet 5.5 on 28 September at the same list price as Sonnet 5, 2 dollars per million input tokens and 10 dollars per million output tokens, half the 4 and 20 dollars charged for Opus 5.5.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source On Anthropic's own table the mid tier model now sits within a few points of the flagship on most evaluations and ahead of it on Terminal-Bench 4.0, an agentic coding test, at 70.6 percent against 66.4 percent.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source The commercial consequence lands first inside Anthropic: a customer who can get near flagship results for half the price has little reason to pay for the flagship on routine work. We assess with high confidence that Sonnet 5.5 will take a large share of Opus 5.5 usage on the Claude Platform, and with moderate confidence that the launch tightens price pressure on OpenAI's mid tier rather than on its frontier model.

What the numbers say, and what they do not

The headline jump is on Terminal-Bench 4.0, where Sonnet 5 scored 10.3 percent.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source A sevenfold improvement in one generation on a single benchmark is unusual enough to read carefully. The test measures how well an agent completes multi step engineering work in a live command line, and it rewards models that recover from tool errors rather than models that write better code in isolation. Gains of that size on a new benchmark version often reflect a model learning the harness as much as the task.

The broader table is the more useful signal. On GDPval-AA v2.1, a professional work evaluation scored as an Elo style rating, Sonnet 5.5 posts 1844 against 1846 for Opus 5.5 and 1449 for Sonnet 5.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source On OSWorld 2.1, a computer use test, it scores 80.1 percent against 81.8.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source On FrontierCode 1.1 and Humanity's Last Exam with tools it trails the flagship by several points.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source The honest summary is that Sonnet 5.5 is roughly an Opus 5.5 for work that does not require the hardest reasoning, at half the price, and somewhat faster: Anthropic claims output more than 30 percent faster and per task cost lower by up to 30 percent for most work.2 Android Headlines 2026-09-28 Availability on AWS, Google Cloud and Azure; over 30 percent faster; up to 30 percent cheaper per task; Max effort caused sub agent timeouts; GPT-6 Sol at 1,487 on GDPval-AA. Open source

The caveats are real. Every figure is vendor reported, and scores are quoted at different effort settings across models, so cross model comparisons are not like for like.3 Vellum 2026-09-29 Vendor sourced data; inconsistent effort settings; Max scored 46.2 percent versus 52.1 at Xhigh on FrontierCode; GPT-6 Sol missing from several benchmarks; customer claims unaudited. Open source On FrontierCode, Sonnet 5.5 scored lower at its Max effort setting than at Xhigh, 46.2 percent against 52.1, and Android Headlines reports that Max triggered extra sub agent reviews that caused timeouts.3 Vellum 2026-09-29 Vendor sourced data; inconsistent effort settings; Max scored 46.2 percent versus 52.1 at Xhigh on FrontierCode; GPT-6 Sol missing from several benchmarks; customer claims unaudited. Open source 2 Android Headlines 2026-09-28 Availability on AWS, Google Cloud and Azure; over 30 percent faster; up to 30 percent cheaper per task; Max effort caused sub agent timeouts; GPT-6 Sol at 1,487 on GDPval-AA. Open source OpenAI's GPT-6 Sol is missing from several rows.3 Vellum 2026-09-29 Vendor sourced data; inconsistent effort settings; Max scored 46.2 percent versus 52.1 at Xhigh on FrontierCode; GPT-6 Sol missing from several benchmarks; customer claims unaudited. Open source Independent evaluation will narrow or widen these gaps within weeks.

The safeguards are part of the product

Anthropic describes Sonnet 5.5 as shipping with cyber safeguards similar to Opus 5.5: routine bug fixing is allowed, while higher risk security tasks fall back to Sonnet 5, and an expanded verification program offers tiered access to vetted defenders.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source It also ships new classifiers to block reasoning extraction, with preserved thinking tied to the account that created it.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source That second measure protects the model's reasoning from distillation, and it imposes a cost on legitimate users who move conversations between accounts.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source

The fallback design is worth noting because it means the model a customer pays for is not always the model that answers. For most buyers this is invisible. For security teams it is a capability ceiling enforced by a classifier, and the verification program is the gate around it.

Who gains and who loses

Developers building agents gain the most. A model that scores close to the flagship on computer use and professional work at half the price makes agent products cheaper to run at volume, and availability on Amazon Web Services, Google Cloud and Microsoft Azure from day one removes procurement friction.2 Android Headlines 2026-09-28 Availability on AWS, Google Cloud and Azure; over 30 percent faster; up to 30 percent cheaper per task; Max effort caused sub agent timeouts; GPT-6 Sol at 1,487 on GDPval-AA. Open source

Opus 5.5 loses share by design. Anthropic is choosing to cannibalise its own top tier, which only makes sense if it expects revenue from volume growth to exceed the margin given up per request. OpenAI's GPT-6 Sol is the external comparison on the vendor's own table, and it trails Sonnet 5.5 on GDPval-AA at 1487.2 Android Headlines 2026-09-28 Availability on AWS, Google Cloud and Azure; over 30 percent faster; up to 30 percent cheaper per task; Max effort caused sub agent timeouts; GPT-6 Sol at 1,487 on GDPval-AA. Open source If those numbers survive independent testing, OpenAI faces pressure to answer on price at the mid tier.

Cyber researchers outside the verification program lose some capability relative to the new model, since high risk requests are routed to the older one.1 Anthropic 2026-09-28 Released 28 Sep 2026; 2 and 10 dollars per million tokens versus 4 and 20 for Opus 5.5; Terminal-Bench 4.0 70.6 percent versus 10.3 for Sonnet 5 and 66.4 for Opus 5.5; GDPval-AA 1844 versus 1846; OSWorld 2.1 80.1 versus 81.8; cyber fallback to Sonnet 5; anti distillation classifiers; Haiku 5.5 in coming weeks. Open source

The counter case

The main assessment could be wrong if the benchmark gains do not transfer. Vendor tables are tuned to show a model at its best effort setting, and the customer quotes behind the launch are unaudited.3 Vellum 2026-09-29 Vendor sourced data; inconsistent effort settings; Max scored 46.2 percent versus 52.1 at Xhigh on FrontierCode; GPT-6 Sol missing from several benchmarks; customer claims unaudited. Open source If developers find that Sonnet 5.5 fails more often on long, messy, real repositories than Opus 5.5 does, the flagship keeps its premium and the cannibalisation never happens. The FrontierCode regression at Max effort is a small hint that the model's behaviour at its highest settings is not yet stable.3 Vellum 2026-09-29 Vendor sourced data; inconsistent effort settings; Max scored 46.2 percent versus 52.1 at Xhigh on FrontierCode; GPT-6 Sol missing from several benchmarks; customer claims unaudited. Open source

What to watch

  • Haiku 5.5 lands cheap, and soon. Our expectation is a release inside two weeks, priced aggressively, perhaps a tenth of a dollar per million input tokens. It would likely come alongside a further cut to Sonnet 5.5 cache pricing. Anthropic is moving the whole family down the cost curve, not defending a mid tier floor.

What happened next (updated Oct 13, 2026)

Anthropic released Claude Haiku 5.5 on 7 October at 0.10 dollars input and 0.50 dollars output per million tokens for prompts up to 100,000 tokens, and at the same time cut Sonnet 5.5 cache reads from 0.20 to 0.10 dollars per million tokens.4 Let's Data Science 2026-10-07 Haiku 5.5 released 7 Oct 2026; 0.10 and 0.50 dollars per million tokens under 100k tokens; Sonnet 5.5 cache reads cut to 0.10 dollars. Open source The family moved down the cost curve together rather than holding the mid tier price as a floor.