Anthropic kept its word. When Claude Opus 5.5 launched on September 22 promising Sonnet 5.5 and Haiku 5.5 'within weeks,' it took six days for the first half of that promise to arrive. Claude Sonnet 5.5 launched September 28, 2026, and the standout number isn't a new capability ceiling — it's that Anthropic's mid-tier model now posts a Terminal-Bench 4.0 score nearly seven times higher than its immediate predecessor, at the exact same price.
That's a genuinely unusual jump for a same-tier, same-price refresh, and it deserves the caveat that comes with any vendor-reported number this large: it's measured at Anthropic's maximum reasoning effort setting, which costs real money per task. Here's what Sonnet 5.5 actually ships with, what the number means in practice, and how it stacks up against Anthropic's own more expensive models.
Quick summary: Anthropic released Claude Sonnet 5.5 on September 28, 2026 (model ID claude-sonnet-5-5), live on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. Pricing is unchanged from Sonnet 5 at $2 per million input tokens and $10 per million output tokens. Anthropic says it's over 30% faster and up to 30% cheaper for typical work, with a 1-million-token context window and a 128K-token maximum output. On Terminal-Bench 4.0 — a test of agentic coding inside a command-line terminal — Anthropic reports Sonnet 5.5 scoring 70.6% at maximum reasoning effort, against 10.3% for Sonnet 5 on the same benchmark.
The Terminal-Bench number, with the caveat attached
The 70.6% vs. 10.3% figure is the headline most coverage led with, and it's real — but it comes from Sonnet 5.5's highest effort setting, where a single attempt reportedly costs more than six times what it does at the API's default effort level. At Medium effort, the setting used by default in the Claude apps, Anthropic says Sonnet 5.5 still comfortably exceeds Sonnet 5's best score, but for a fraction of the cost per task. In other words: the eye-catching 7x number is real but expensive to actually reproduce; the practical, everyday improvement is smaller in absolute terms but dramatically cheaper to get (Sources: apidog).
How it stacks up against Anthropic's own pricier models
Anthropic now sells four public tiers in the Claude 5-generation lineup: Haiku, Sonnet 5.5, Opus 5.5, and Fable 5.1 (its actual top public tier, with the gated Mythos 5.1 above that). On the independent Artificial Analysis Intelligence Index — a composite benchmark run by a third party, not Anthropic itself — Sonnet 5.5 reportedly scores 56, against 58 for Opus 5.5 and 53 for Fable 5.1. That's a striking result on its face: a model priced at $2/$10 per million tokens scoring within a few points of a flagship-adjacent model priced at $4/$20, and ahead of Anthropic's actual top public tier at $10/$50. On Terminal-Bench 4.0 specifically, Sonnet 5.5's reported 70.6% even nominally edges out Opus 5.5's own reported 66.4% — though Anthropic's published error margins (roughly ±2.5 to ±2.6 points on this 66-task benchmark) mean that particular gap isn't a clear win either way.
| Model | Input / 1M tokens | Output / 1M tokens | Terminal-Bench 4.0 (Anthropic, max effort) |
|---|---|---|---|
| Claude Sonnet 5 | $2 | $10 | 10.3% |
| Claude Sonnet 5.5 | $2 | $10 | 70.6% |
| Claude Opus 5.5 | $4 | $20 | 66.4% |
| Claude Fable 5.1 | $10 | $50 | not published |
That comparison comes with its own honesty check: Opus 5.5's own Terminal-Bench number has already been questioned once this month. Independent testing cited by multiple outlets put Opus 5.5 closer to 59.6% on the same benchmark, roughly level with GPT-6 Astra rather than clearly ahead — a reminder that Anthropic's own charts and third-party reruns don't always agree, on Sonnet 5.5 or anything else, until outside labs catch up with their own numbers.
What's actually new beyond the benchmark
- Context window: 1 million tokens, with a maximum output of 128,000 tokens per response — useful for long agentic sessions or large-document work.
- Speed: Anthropic says output generation is more than 30% faster than Sonnet 5.
- Cost for typical work: Anthropic frames the overall cost reduction for everyday tasks at up to 30%, separate from the unchanged per-token list price — the saving comes from doing more per dollar, not a lower sticker price.
- Availability: live now via the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
- Haiku 5.5 is still pending; Anthropic hasn't given it a firm date beyond the original 'within weeks' of Opus 5.5.
If you're deciding between Sonnet 5.5 and Opus 5.5 for agentic coding work, run your own task at both models' default (not maximum) effort settings before paying for the pricier tier — on Anthropic's own numbers, the gap between them narrows a lot once you're not comparing best-case, most-expensive runs.
01Is Claude Sonnet 5.5 actually better than Claude Opus 5.5?
On a few specific benchmarks at maximum effort, Anthropic's own numbers put Sonnet 5.5 nominally ahead or roughly even — for instance, 70.6% vs. 66.4% on Terminal-Bench 4.0. But Opus 5.5 still scores higher on the independent Artificial Analysis Intelligence Index (58 vs. 56), and Anthropic's own error margins mean a single benchmark lead shouldn't be read as Sonnet 5.5 broadly outclassing Opus 5.5.
02Did the price change with Sonnet 5.5?
No. It's priced identically to Sonnet 5 at $2 per million input tokens and $10 per million output tokens. The cost savings Anthropic advertises come from efficiency (needing fewer tokens or less time per task), not a lower per-token rate.
03What does the 70.6% vs 10.3% Terminal-Bench number actually mean?
Terminal-Bench 4.0 tests how well a model completes multi-step technical tasks inside a command-line environment. Both scores are at Anthropic's maximum reasoning effort setting, which is also the most expensive setting to run — the everyday, default-effort improvement is smaller but far cheaper per task.
04Is Haiku 5.5 out yet?
Not as of this post. Anthropic said when it released Opus 5.5 on September 22 that Sonnet 5.5 and Haiku 5.5 would both follow within weeks; Sonnet 5.5 arrived six days later, but Anthropic hasn't given Haiku 5.5 a specific date.
05Where can I use Claude Sonnet 5.5?
It's available now through the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
Benchmark tables only tell you so much about whether a model fits your own work. LumiChats lets you chat with Claude Sonnet 5.5 and 40+ other leading AI models side by side in one place, so you can test a model against your actual prompts before deciding which one earns a place in your workflow.
