Anthropic Releases Claude Sonnet 5.5, Nearly Matching Opus 5.5 at Up to 30% Lower Cost Per Task

Anthropic has released Claude Sonnet 5.5, the second model in its Claude 5.5 family, generating output more than 30 percent faster and cutting per-task costs by up to 30 percent through more efficient token usage. The model nearly matches flagship Opus 5.5 on several benchmarks at a fraction of the compute cost, according to Anthropic.

Sonnet 5.5 targets well-defined everyday work — bug fixes, documentation, presentations, spreadsheets — while Opus 5.5 remains positioned for complex tasks requiring careful judgment. Anthropic also announced Claude Haiku 5.5 for "the coming weeks," aimed at high-throughput, low-cost use cases.

Coding gains lead the improvements

The largest jump between Sonnet 5.5 and its predecessor appears in coding benchmarks, per Anthropic's released data:

• Terminal-Bench 4.0 (agentic coding): 70.6% vs. Sonnet 5's 10.3% — ahead of Opus 5.5's 66.4%
• CursorBench 4.0 (real Cursor editor sessions): 55.5% vs. 34.1%, trailing Opus 5.5 by just 2.3 points (57.8%)
• FrontierCode 1.1 (High setting): roughly 10 points above Sonnet 5 at about one-fifteenth the cost per task, Anthropic claims

One anomaly: at the highest "Max" effort setting, Sonnet 5.5 scores worse on FrontierCode (46.2%) than at "Xhigh" (52.1%). Anthropic attributes this to the model more frequently triggering a code-review function that splits work across sub-agents, sometimes causing timeouts or out-of-scope changes that the benchmark penalizes.

Knowledge work approaches Opus 5.5

On GDPval-AA v2.1, an OpenAI-developed benchmark spanning 44 professions across nine industries, Sonnet 5.5 scores 1,844 points — nearly matching Opus 5.5's 1,846 and roughly 400 points above Sonnet 5's 1,449. Anthropic's figures put OpenAI's GPT-6 Sol at 1,487 on the same test. On AA Briefcase v1.1, Sonnet 5.5 hits 1,811 versus Opus 5.5's 1,822.

On Chartography, a visual chart-recognition test, Sonnet 5.5 jumps from 15.6% to 61.6%, ahead of GPT-6 Sol's 53.6%. On Humanity's Last Exam with tools, it scores 64.5% versus Opus 5.5's 67.7%. On OSWorld 2.1 (partial evaluation), it reaches 80.1% against Opus 5.5's 81.8%.

Anthropic notes that some GDPval-AA and AA Briefcase figures for Sonnet 5.5 came from a pre-release build where a now-fixed bug may have affected structured outputs.

Pricing unchanged, but fewer tokens needed

Sonnet 5.5 keeps Sonnet 5's per-token pricing: $2 per million input tokens, $10 per million output tokens, $0.20 for cache reads, and $2.50 for cache writes. Anthropic says the effective cost drop comes from the model needing fewer tokens per task, not a price cut. Output generation speed is also up more than 30 percent, according to the company. These efficiency claims have not yet been independently verified.

Availability and new safeguards

Sonnet 5.5 is live now on AWS, Google Cloud, and Microsoft Azure, accessible via the Claude Platform under the model ID claude-sonnet-5-5, with zero data retention offered as an option.

Because the model shows substantially stronger cybersecurity capability than its predecessor, Anthropic is applying safeguards to a Sonnet model for the first time: high-risk cybersecurity requests are automatically rerouted to Sonnet 5, and an expanded Cyber Verification Program allows vetted professionals tiered access. Anthropic has also added distillation-attack classifiers, previously reserved for its most capable models.

What this means

Sonnet 5.5 narrows the performance gap with Opus 5.5 to single digits on several agentic and knowledge-work benchmarks while claiming lower effective cost — a pattern that mirrors OpenAI's GPT-6 tier structure (Astra/Sol/Luna) and signals that mid-tier models are absorbing capabilities once reserved for flagship models. With per-tier performance converging across labs, pricing and token efficiency — not raw benchmark scores — may increasingly decide which model developers default to. The new distillation-attack classifiers are also notable: if they prove effective, they could meaningfully slow how quickly rival labs, including Chinese developers who have reportedly used distillation to catch up, can close the gap on frontier capability.