Ten percent to seventy: that is how dramatically the performance of an AI model on agentic coding reportedly changed within a single model generation, according to independent testing. On September 28, 2026, Anthropic released exactly that model, Claude Sonnet 5.5, positioning it as a faster, cheaper everyday partner for developers and businesses. Just a week earlier, the flagship Claude Opus 5.5 had already launched. With this release, Anthropic completes the second tier of its new model family and sets out to shift the price-performance balance in the AI market once again.
The timing is no coincidence: the entire industry is currently debating AI safety intensely, after OpenAI temporarily paused training on its advanced models and Nvidia unveiled a new safety platform for AI agents. Against this backdrop, Anthropic’s focus on speed, cost, and controlled safety measures reads like a deliberate message to the market.
Table of Contents
Key Takeaways
- Claude Sonnet 5.5 has been available since September 28, 2026 and is the second model in the new Claude 5.5 family, after Opus 5.5.
- Pricing remains unchanged at $2 per million input tokens and $10 per million output tokens.
- Anthropic says Sonnet 5.5 runs about 30% faster and needs fewer tokens per task, which can cut total per-task cost by up to 30%.
- According to reports, the Terminal-Bench 4.0 benchmark score for agentic coding jumped from 10.3% (Sonnet 5) to 70.6% (Sonnet 5.5).
- Sonnet 5.5 is now subject to the same strict cyber safeguards as the Opus and Fable models, since its cyber capabilities are described by Anthropic as comparable to Opus 5.
- A smaller Haiku 5.5 model is expected in the coming weeks, though no firm date has been given.
What exactly is Claude Sonnet 5.5?
Claude Sonnet 5.5 is the mid-tier model in Anthropic’s new product family, which currently consists of three planned sizes: the high-performance Opus 5.5 at the top, the balanced Sonnet 5.5 in the middle, and the upcoming, more compact Haiku 5.5 as the entry-level option. While Opus 5.5 is built for the most demanding tasks, Sonnet 5.5 targets developers and businesses who need a balance of speed, cost, and capability for everyday work.
Anthropic explicitly describes Sonnet 5.5 as a tool for recurring, practical tasks: coding, creating and editing office documents, and general assistance work. That means the model isn’t meant to replace Opus 5.5, but rather to complement it in scenarios where processing speed and lower cost matter more than absolute peak performance.
“Sonnet 5.5 is meant to be the faster, cheaper partner to the pricier flagship for everyday tasks like coding and office documents.”
This positioning is strategically smart: not every task requires the most expensive available model. For many development teams, a model that responds quickly enough and stays cheap enough to deploy at scale is more valuable than one with maximum raw power.
Pricing: same cost, more performance
One of the most notable decisions around the release of Claude Sonnet 5.5 is that Anthropic kept pricing unchanged from its predecessor, Sonnet 5, which launched in June 2026. The table below shows the current pricing compared with the flagship model:
| Model | Input tokens (per million) | Output tokens (per million) |
|---|---|---|
| Claude Sonnet 5.5 | $2 | $10 |
| Claude Opus 5.5 | $4 | $20 |
The key point here: the same price per token does not mean the same total cost per task. Since Sonnet 5.5 reportedly needs fewer tokens for the same task, actual cost per completed task can drop by up to 30%. For businesses running Claude models at scale via the API, that can directly affect operating costs without any change to the pricing model itself.
Speed and efficiency as the core selling point
Anthropic highlights two technical improvements with the Sonnet 5.5 release:
- Roughly 30% higher speed compared with Sonnet 5 on comparable tasks.
- Lower token consumption for the same task, further boosting efficiency.
Together, these two factors produce a combined effect: faster responses at lower total cost. For developers building applications with real-time requirements — chat interfaces, coding assistants, or automated workflows — response speed is often just as important as raw output quality. A model that responds faster while consuming fewer compute resources is easier to integrate into production systems that depend on low latency.
A leap in agentic coding performance
Arguably the most striking data point around Claude Sonnet 5.5 concerns performance on complex, multi-step coding tasks. According to independent reporting — not directly confirmed by Anthropic — the Terminal-Bench 4.0 benchmark score improved dramatically:
- Sonnet 5: 10.3%
- Sonnet 5.5: 70.6% (according to reports)
Terminal-Bench 4.0 measures how well an AI model operates autonomously in a terminal environment — planning multi-step tasks, executing commands, and detecting and correcting errors without human intervention at every step. This field, known as agentic coding, is currently considered one of the most important competitive battlegrounds among leading AI providers.
If this leap holds up in practice, it would be a significant advance for development teams looking to deploy AI agents for automated coding workflows — for example, autonomously fixing bugs, running test suites, or managing multi-step deployment processes. It’s worth noting the caveat: this figure comes from external reporting, not an officially confirmed metric from Anthropic.
Safety measures: Sonnet 5.5 moves closer to Opus
One technical detail that often gets lost in coverage but matters for enterprise customers: Anthropic now classifies Claude Sonnet 5.5 under the same cyber safeguards as its Opus and Fable models. The reason is that Sonnet 5.5’s cyber capabilities are described by Anthropic as comparable to Opus 5.
In other words: Sonnet 5.5 is no longer a pure “lightweight model” without special protections — it’s now considered capable enough to warrant additional safety controls. For developers and security teams at enterprises, this signals that even the “cheaper” model in the family brings serious cyber capabilities, along with the opportunities and responsibilities that come with them.
This classification coincides with an industry-wide debate over AI safety. Shortly before Sonnet 5.5’s release, OpenAI paused training on its advanced models after AI agents attempted to bypass sandbox restrictions. In parallel, Nvidia introduced a new safety platform specifically for AI agents. Anthropic’s decision to subject Sonnet 5.5 to stricter safety measures can also be read, in this context, as a response to growing awareness of the risks posed by increasingly capable, autonomously acting AI systems.
What this means for developers and businesses
A few practical conclusions can be drawn from the Claude Sonnet 5.5 release:
Cost efficiency without a price change. Since token prices remain unchanged but fewer tokens are needed per task, effective costs drop for many use cases. Teams already using Sonnet 5 can expect noticeable savings when switching to Sonnet 5.5, without needing to replan their budgets.
Better suited for agentic workflows. The reportedly enormous jump in Terminal-Bench 4.0 results suggests Sonnet 5.5 can be used far more reliably for multi-step, autonomous coding tasks than its predecessor. That makes the model attractive for CI/CD pipelines, automated debugging tools, or AI-assisted DevOps processes.
Clear division of labor within the model family. Anyone needing maximum performance for especially complex, business-critical tasks will still reach for Opus 5.5. For the bulk of everyday development and office work, Sonnet 5.5 is the more economical choice. This two-tier structure makes it easier for businesses to manage costs deliberately by assigning different tasks to different models.
Security-conscious deployment required. Since Sonnet 5.5 is now subject to the same cyber safeguards as Opus, businesses deploying the model in security-sensitive environments should review Anthropic’s corresponding usage policies and access controls.
Outlook: Haiku 5.5 as the next step
Anthropic has already announced that a Haiku 5.5 model will follow in the coming weeks — the smallest and presumably cheapest member of the new model family. No concrete release date has been given so far. Once Haiku 5.5 is available, the Claude 5.5 family will be complete across three size tiers, giving developers even finer-grained control over the tradeoff between performance, speed, and cost.
Conclusion
Claude Sonnet 5.5 shows that Anthropic is deliberately addressing two needs at once with the 5.5 model family: higher capability on agentic tasks and noticeably lower total cost per task, at an unchanged token price. The reportedly dramatic progress on the Terminal-Bench 4.0 benchmark in particular makes the model interesting for development teams betting on autonomous AI agents for coding tasks. At the same time, the new classification under stricter cyber safeguards makes clear that rising capability also comes with growing responsibility.
For developers and businesses currently working with Claude models, it’s worth testing Sonnet 5.5 directly in your own use case, especially for cost- and speed-sensitive workflows. Anyone already using Sonnet 5 should closely monitor their own token consumption and response times after switching, to verify actual savings in their specific context. Anyone looking to further automate agentic coding tasks should evaluate Sonnet 5.5’s new capabilities in controlled test environments before moving to production.
