Anthropic's Claude Sonnet 5.5 Cuts Per-Task Costs by 30% With No Price Increase

30-Second Summary

Anthropic released Claude Sonnet 5.5 on 28 September 2026. It generates output more than 30% faster than Sonnet 5 and, because it completes tasks in fewer steps, reduces real-world cost per task by up to 30%. Token pricing is unchanged. The model's agentic coding score on Terminal-Bench 4.0 jumped from 10.3% to 70.6%, and on knowledge-work benchmarks it nearly matches Opus 5.5 while costing roughly half as much. Enterprise teams already using Claude via API or Claude.ai receive this improvement automatically.

At a Glance

• Topic: Model release
• Company: Anthropic
• Date: 28 September 2026
• Announcement: Claude Sonnet 5.5 released with 30%+ speed improvement and up to 30% lower per-task cost
• What Changed: The model completes tasks in fewer steps and generates tokens faster, delivering cost savings without a price cut
• Why It Matters: The default mid-tier Claude model most enterprise teams use day-to-day became substantially more capable and cost-effective overnight
• Who Should Care: Any organisation using Claude API, Claude Teams, or Claude Enterprise, and any team evaluating Claude against GPT-6 or Gemini 4

Key Facts

• Token pricing: $2 per million input tokens, $10 per million output tokens, $0.20 per million cache-read tokens (same as Sonnet 5)
• Speed: output generates more than 30% faster than Sonnet 5
• Cost per task: up to 30% lower due to fewer tool calls required to complete the same task
• Terminal-Bench 4.0 (autonomous command-line agentic coding): 70.6% versus Sonnet 5's 10.3%
• CursorBench 4.0 (IDE-based coding): 55.5%, second only to Opus 5.5
• GDPval-AA v2.1 (professional knowledge work): 1,844 Elo, two points shy of Opus 5.5 and 360 Elo points ahead of GPT-6 Sol
• Box production result: Claude runs complete 2.4x faster (Source: Anthropic, September 2026)
• Zendesk production result: tickets processed 20% faster (Source: Anthropic, September 2026)

What Happened

Anthropic published Claude Sonnet 5.5 on 28 September 2026. The model sits in the same position in Anthropic's lineup as previous Sonnet versions: the mid-tier workhorse, positioned below Opus for raw capability but above Haiku for task complexity. It is the model most enterprise teams use for production deployments.

The key engineering improvement is not simply token generation speed, though that is measurably faster. Sonnet 5.5 completes the same tasks in fewer steps. In agentic workflows, where a model might call ten tools to answer a question or complete a process, the new model typically calls seven or eight. Those saved steps compound into the 30% per-task cost reduction Anthropic reports.

The most striking benchmark result is on Terminal-Bench 4.0, which measures how well an autonomous AI agent completes complex engineering tasks inside a live command-line environment without human help. Sonnet 5 scored 10.3% on this benchmark. Sonnet 5.5 scores 70.6%, outscoring even Opus 5.5's top mark. That is not an incremental improvement: it reflects a step-change in autonomous task completion reliability.

Enterprise customers Box and Zendesk both provided production results. Box reported their Claude API integrations now complete runs 2.4 times faster. Zendesk reported support tickets processed 20% faster. Both companies were early access testers (Source: Anthropic, September 2026).

Why It Matters

The default enterprise model just improved substantially for free. Most teams using Claude Enterprise or the Claude API are running Sonnet by default. They did not need to migrate to a new model, adjust their prompts, or change their billing. The improvement arrived as an automatic update.

Agentic workflows become substantially more reliable. The jump from 10.3% to 70.6% on Terminal-Bench 4.0 is the signal that autonomous, multi-step tasks are crossing a reliability threshold. Tasks that previously needed a human to review step results every few minutes can now complete without intervention. This is the change that makes AI automation practical across a wider range of internal workflows.

Sonnet 5.5 now competes with Opus for most tasks. On knowledge-work benchmarks, the two models are nearly indistinguishable. On agentic coding, Sonnet 5.5 outperforms Opus 5.5 outright. For teams that were paying Opus prices to get sufficient quality, the same budget now stretches roughly twice as far.

The cost reduction compounds across large-scale deployments. A team running 10,000 agentic task completions per month at Sonnet pricing already has a low base rate. A further 30% reduction per task, at no pricing change, meaningfully affects the economics of deploying AI across multiple departments simultaneously.

The enterprise AI cost argument just got easier to make. Procurement teams and CFOs evaluating Claude deployments typically focus on total cost at scale. Sonnet 5.5's efficiency improvement directly addresses the most common objection: that token costs multiply unpredictably as usage grows.

The David and Goliath View

The Sonnet tier has always been the practical choice for Claude Enterprise deployments. Opus is positioned as the most capable model but costs more per token; Haiku handles simple queries at speed. Sonnet is where most real workflows live, and Sonnet 5.5 is a meaningful upgrade to the model that does the actual work in most businesses.

What the Terminal-Bench 4.0 result signals is that the gap between a capable language model and a reliable autonomous agent is closing. A model that scores 70.6% on autonomous command-line engineering tasks is useful enough to run unattended on a large proportion of standard development and operations tasks. That is a different proposition than a model that needs supervision every few steps.

For organisations in the middle of a Claude Activation Sprint, or evaluating whether to deploy Claude across finance, legal, or operations functions, the timing is helpful: the standard model just became substantially better without requiring any migration or renegotiation.

Where This Fits in the AI Stack

Sonnet 5.5 sits at the centre of the Claude model stack, which runs from Claude Haiku 4.5 for speed-critical lightweight tasks up through Claude Opus 5.5 for the most demanding reasoning work. It is the model most Claude Enterprise API integrations are pointed at by default. The model is available now through the Anthropic API, Claude.ai Teams, and Claude Enterprise. Prompt caching (storing repeated context to reduce input token costs across long sessions) is available at standard rates.

Questions Operators Are Asking

Do we need to update our API calls or system prompts to get the improvement? No. Existing API calls pointing to claude-sonnet-5 will continue to work. To access Sonnet 5.5 specifically, update your model parameter to claude-sonnet-5-5. Anthropic has not announced a date for Sonnet 5 deprecation.

Is the 30% cost reduction per task or per token? Per task. The token price is identical to Sonnet 5. The cost reduction comes from the model using fewer tool calls to complete the same amount of work. Individual token costs are unchanged; the saving accumulates across multi-step workflows.

Should we migrate from Opus 5.5 to Sonnet 5.5? For coding tasks and most knowledge work, yes. Sonnet 5.5 outperforms Opus 5.5 on Terminal-Bench 4.0 and matches it on GDPval-AA v2.1. Tasks that specifically require Opus-level reasoning on complex, multi-part problems should stay on Opus. A practical approach is to run Sonnet 5.5 for the bulk of production traffic and keep Opus available for the subset of tasks that exceed Sonnet's capabilities.

How does it compare to GPT-6 Sol for enterprise knowledge work? On GDPval-AA v2.1, Sonnet 5.5 scores 1,844 Elo compared to GPT-6 Sol's result roughly 360 points lower. On enterprise knowledge work benchmarks, Sonnet 5.5 is the current leader among mid-tier models.

What changed in the model's architecture to produce this improvement? Anthropic has not published architectural details. The observable effect is that the model completes tasks with fewer intermediate steps, which suggests improved planning and task decomposition. The agentic coding jump on Terminal-Bench 4.0 is the clearest signal of this.

Citable Summary

Anthropic released Claude Sonnet 5.5 on 28 September 2026. The model generates output more than 30% faster than Sonnet 5 and reduces real-world cost per task by up to 30% through improved efficiency, not a price cut. Token pricing is unchanged at $2 per million input tokens. On Terminal-Bench 4.0, an autonomous agentic coding benchmark, the model scores 70.6% versus Sonnet 5's 10.3%. On professional knowledge-work benchmarks, it nearly matches Opus 5.5 while costing roughly half as much. Enterprise customers Box and Zendesk reported production speed improvements of 2.4x and 20% respectively. (Source: Anthropic, September 2026.)