The real cost of AI can no longer be read off a price sheet. On September 28, 2026, Anthropic released Claude Sonnet 5.5 at the same price as its predecessor, 2 dollars per million input tokens, 10 dollars per million output tokens, while announcing a bill up to 30% lighter for the same task. For an SME leader comparing AI offers on a pricing grid, this release exposes a blind spot: price per token is only part of the equation.
In brief
- Anthropic launched Claude Sonnet 5.5 on September 28, 2026, at the same price as Sonnet 5 ($2 per million input tokens, $10 per million output tokens).
- The model generates responses 30% faster and uses far fewer tokens to reach an equivalent result, cutting the bill for a given task by up to 30%.
- On the agentic benchmark Terminal-Bench 4.0, the score jumps from 10.3% (Sonnet 5) to 70.6% (Sonnet 5.5), ahead of Claude Opus 5.5 (66.4%), Anthropic's most capable model.
- Customers cited by Anthropic, such as Box, report processing 2.4 times faster while consuming 12% fewer tokens.
- The lesson for SMEs goes beyond this single model: comparing AI tools on sticker price per token alone is a flawed analysis.
What Claude Sonnet 5.5 actually changes
Claude Sonnet 5.5 is the mid-tier model in the Claude lineup, sitting between the fast, lightweight model (Haiku) and the most capable one (Opus 5.5). Anthropic positions it for everyday business tasks: fixing code, drafting documents, building presentations and spreadsheets.
Three concrete claims accompany the launch:
- Speed: responses come out more than 30% faster than Sonnet 5, at comparable quality.
- Token efficiency: the model needs fewer back-and-forth steps and less generated text to reach the same result, which lowers the real bill even though the unit price stays the same.
- Stronger safeguards: Sonnet 5.5 is the first Sonnet model subject to the same cybersecurity safeguards as the Opus models, with an automatic fallback to Sonnet 5 on tasks flagged as most sensitive.
Quick definition
Terminal-Bench measures an AI agent's ability to complete real technical command-line tasks on its own: navigating code, fixing a bug, running scripts. A high score signals a model that can be trusted with technical automation without constant human supervision.
The metric that actually matters: cost per task
This is the central point for an SME. Anthropic states that Sonnet 5.5 handles the same missions as Sonnet 5 while consuming noticeably fewer tokens. Two models priced identically per token can therefore produce very different bills depending on their real efficiency.
The Terminal-Bench 4.0 test illustrates the gap starkly:
A jump from 10.3% to 70.6% on agentic terminal tasks within a single model generation means, in practice, fewer failures to fix by hand, and therefore less human time billed behind the automation.
Sonnet 5.5 versus Opus 5.5: which model for your SME
Sonnet 5.5 does not replace Opus 5.5 across the board. On pure reasoning or highly complex tasks (FrontierCode, OSWorld), Opus 5.5 still leads. The real difference lies in the balance between performance and budget for everyday work.
Claude Opus 5.5
Claude Sonnet 5.5
| Criterion | Claude Sonnet 5.5 | Claude Opus 5.5 |
|---|---|---|
| Input / output price (per million tokens) | $2 / $10 | $4 / $20 |
| Terminal-Bench 4.0 (agentic) | 70.6% | 66.4% |
| Recommended use | Volume, everyday tasks | Complex reasoning, lower volume |
| Safety fallback on sensitive tasks | To Sonnet 5 | N/A |
Source: Anthropic, Claude Sonnet 5.5 launch page, September 28, 2026.
Why this release fits a broader trend
This launch confirms a movement already visible the week before, when Anthropic and OpenAI both cut prices on the same day. But Sonnet 5.5 adds an important nuance: the drop in real cost no longer runs only through the sticker price, it also runs through how efficiently a model completes a task with fewer resources. For an SME, this means it is now worth measuring actual usage (tokens consumed per typical task) rather than comparing pricing grids on paper.
FAQ
Is Claude Sonnet 5.5 available everywhere?
Yes. Anthropic made it available starting September 28, 2026 on its Claude platform, as well as on Amazon Web Services, Google Cloud and Microsoft Azure, under the model ID claude-sonnet-5-5.
Should my SME switch models right away?
Not automatically. For an SME already using Sonnet 5, swapping the model name is usually enough to switch over, but it is better to test on a real use case before rolling it out broadly, to confirm quality holds up on your own documents or code.
How do I know if an AI model is actually cheaper for my business?
Price per million tokens is not enough on its own. You need to measure how many tokens a typical task in your company consumes (a support ticket, a report generation) and compare the final cost, not the catalog price.
What is Terminal-Bench and why does it matter?
It is a test that evaluates an AI agent's ability to complete technical command-line tasks on its own. A high score indicates a model more reliable for technical automation without constant human intervention, a useful criterion if your SME is considering AI agents for support or development work.
Is Sonnet 5.5 suitable for a small team with no technical staff?
Yes, since it already powers mainstream tools (writing assistants, document generators) that hide the underlying technical complexity. The real point of attention is choosing the right end tool, not the underlying model.
Choosing an AI model is therefore not just about comparing catalog prices. To go further on how to structure that choice in your SME, see our guide to choosing your AI model or browse our full set of AI resources for business leaders.
