Anthropic Ships Claude Sonnet 5.5 — 30% Faster, 30% Cheaper Than Sonnet 5
Anthropic Ships Claude Sonnet 5.5 — 30% Faster, 30% Cheaper Than Sonnet 5
Anthropic released Claude Sonnet 5.5 on September 28, 2026, the second model in its Claude 5.5 family and the direct successor to Claude Sonnet 5. The headline claims are straightforward: Sonnet 5.5 runs 30% or more faster than Sonnet 5 and costs up to 30% less for most work, with the same per-million-token sticker price as Sonnet 5 ($2 input, $10 output, $0.20 cached input). Real cost savings come from fewer generated tokens and fewer tool calls per task, not from a lower rate card.
Sonnet 5.5 is available today in Claude Code and the Anthropic API, with the model ID claude-sonnet-5-5-20260928 across platforms. It inherits the 200K context window of Sonnet 5 and adds stronger coding safeguards and anti-distillation protections that were previously reserved for Anthropic's flagship models.
The Claude 5.5 family, two tiers
Sonnet 5.5 is the second model in the Claude 5.5 family, following Claude Opus 5.5 (released September 22). The two-tier structure mirrors what OpenAI did in September 2025 with its GPT-6 Sol and Luna pairing, and what AIPress covered at the time as a shift toward "cheaper, broader access" — Post 31 on September 23, 2026.
The pattern is now clear across both frontier labs: the flagship model (Opus 5.5, GPT-6 Astra) handles the hardest tasks where reliability matters, while the mid-tier model (Sonnet 5.5, GPT-6 Sol) handles the high-volume, fast-iteration work where speed and cost per task are the binding constraint.
Opus 5.5 was already covered by AIPress — see Post 32 for the September 22 launch and Post 39 for the September 23 pricing breakdown. Sonnet 5.5 is a distinct product: different model ID, different performance characteristics, and a different target workload (daily coding iteration, not flagship-level reasoning).
The numbers
Anthropic's own benchmark figures for Sonnet 5.5:
| Benchmark | Sonnet 5.5 | Sonnet 5 |
|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 53.8% |
| SWE-Bench Pro | 81.3% | 72.4% |
| CursorBench 4.0 | 55.5% | 41.2% |
| OSWorld 2.1 | 80.1% | 67.9% |
| GDPval-AA v2.1 (Elo) | 1844 | 1720 |
The 30%+ faster claim comes from Anthropic's internal measurement of token generation throughput on identical workloads. The 30% cost reduction is real-world, not sticker-price: fewer output tokens per task (Anthropic reports ~22% shorter outputs on coding tasks) and fewer tool-call rounds (about 18% fewer on average), which compounds to roughly 30% lower effective cost at the same per-token rates.
These are vendor-reported figures, not independent benchmarks. Anthropic's published numbers come from the company's own evaluation harness, and as AIPress noted in its coverage of GPT-6 Astra (Post 1), "vendor-chosen quotes, not independent benchmarks." Third-party verification from Vellum and other evaluation platforms is expected in early October.
What improved
Beyond the headline speed and cost numbers, Anthropic says Sonnet 5.5 introduces three concrete improvements over Sonnet 5:
-
Anti-distillation safeguards — the same protections Anthropic shipped with Opus 5.5, now extended to the mid-tier model. This means Sonnet 5.5 resists efforts to extract its training data more effectively than Sonnet 5 did.
-
Cyber safeguards — Anthropic describes "enhanced tool-use safety" that prevents Sonnet 5.5 from attempting actions outside its authorized scope during coding tasks. These are the same safeguards documented in Post 36 (September 22 coverage of Opus 5.5's cybersecurity features), now ported to the mid-tier.
-
Better coding agent behavior — Anthropic's internal testing shows Sonnet 5.5 requires 22% fewer tool-call rounds to complete multi-step coding tasks compared to Sonnet 5. The model also produces shorter, more targeted outputs, which is where most of the cost savings materialize.
Where it lands
Sonnet 5.5 is available today in:
- Claude Code (Anthropic's IDE extension) — the model is already live
- Anthropic API — model ID
claude-sonnet-5-5-20260928 - Claude.ai — rolling out to Pro and Enterprise users over the next 48 hours
Anthropic has not yet announced distribution on third-party platforms. Post 32 covered Opus 5.5 landing in GitHub Copilot on September 23 — if Sonnet 5.5 follows the same path, that would be a significant distribution win given Copilot's developer base. AIPress is watching the Copilot release notes and Anthropic's blog for that follow-up.
What this means for developers
The Claude 5.5 family now occupies the same pricing position as OpenAI's GPT-6 family:
| Model | Input | Output | Target |
|---|---|---|---|
| Claude Opus 5.5 | $2/M | $10/M | Hard tasks, reliability |
| Claude Sonnet 5.5 | $2/M | $10/M | High-volume coding, iteration |
| GPT-6 Astra | $4.5/M | $18/M | Flagship, agentic |
| GPT-6.1 Sol | $2/M | $10/M | Near-Astra at 1/5 cost |
| GPT-6 Luna | $0.60/M | $2.40/M | Light tasks, cost-sensitive |
Three price points, two vendors, same $2/$10 tier for the mid-tier model. The competition is not just on capability anymore — it is on which lab can deliver the most capable model at the price point developers use every day.
For teams choosing between Claude and OpenAI at the $2/$10 tier, the decision narrows to: do you want Sonnet 5.5's 30% faster throughput and tooling-integrated safety, or GPT-6.1 Sol's "near-Astra" capability at the same price? GPT-6.1 Sol was launched at DevDay 2026 on September 29 — see Post 72 for the full breakdown. The practical answer will depend on which workloads each model benchmarks better on, and third-party evaluations from Vellum and others are expected in October.
The bigger picture
September 2026 is shaping up as the month where "cheaper, broader access" became the dominant narrative across frontier AI — not just a side effect of competition, but the central strategy. Post 33 covered OpenAI and Anthropic dropping new models on the same day (September 23) with the same theme. Sonnet 5.5 is the latest entry in that pattern.
The pricing convergence at $2/M input is notable: it suggests both vendors have settled on a price point where developer adoption crosses a threshold. Below $2/M, the economics pressure the labs' margins. Above $2/M, developers start asking whether the capability uplift justifies the cost. At $2/M, both Sonnet 5.5 and GPT-6.1 Sol are competing flat-out on capability-per-dollar rather than price.
That dynamic favors the developer, at least in the short run. The question for the next quarter is whether either vendor can break the $2/M equilibrium with a meaningfully better model — or whether the market will see another round of price cuts as the labs push toward cost neutrality.
What's not clear yet
- Independent benchmarks: Anthropic's published numbers come from its own harness. Third-party confirmation is expected in October.
- Copilot and IDE distribution: Sonnet 5.5 has not yet been announced for GitHub Copilot or other hosted IDEs. If it follows Opus 5.5's path (covered in Post 32), that is a separate distribution story worth watching.
- Real-world cost savings: The 30% figure is based on Anthropic's internal testing. How it holds up across diverse developer workflows — not just the benchmarks — is not yet publicly verified.
- Safety track record: The anti-distillation and cyber safeguards are improvements over Sonnet 5, but how they hold up under adversarial testing is not yet known.
Named sources
Primary source: Anthropic's announcement post at anthropic.com/claude-sonnet-5-5 (September 28, 2026).
Benchmark sources: Anthropic's model card (September 28, 2026), llm-stats.com model page, Vellum AI blog (expected October 2026).
Coverage: SiliconANGLE (September 28, 2026), 9to5Mac (September 28, 2026), ComputingForgeeks (September 28, 2026), Devex (DevOps.com, September 28, 2026).
Unverified claims: The 30% faster and 30% cost reduction figures are from Anthropic's own testing and have not been independently confirmed as of publication. Copilot/IDE distribution for Sonnet 5.5 has not been announced.
Related AIPress coverage:
- Claude Opus 5.5: Anthropic's New Flagship Matches Fable 5.1 at 40% Less Cost — and Lands in GitHub Copilot — https://aipress.blog/post/claude-opus-5-5-anthropic-s-new-flagship-matches-fable-5-1-at-40-less-cost-and-lands-in-github-copilot
- GPT-6 Sol vs GPT-6 Luna vs Claude Opus 5.5: The September 2026 Model Pricing Breakdown — https://aipress.blog/post/gpt-6-sol-vs-gpt-6-luna-vs-claude-opus-5-5-the-september-2026-model-pricing-breakdown
- OpenAI and Anthropic Dropped New Models on the Same Day — and Both Moved Toward Cheaper, Broader Access — https://aipress.blog/post/openai-and-anthropic-dropped-new-models-on-the-same-day-and-both-moved-toward-cheaper-broader-access
- Anthropic Launches Claude Opus 5.5 with Stricter Cybersecurity Safeguards — https://aipress.blog/post/anthropic-launches-claude-opus-5-5-cybersecurity-safeguards
