Claude Opus vs Sonnet: which tier for each step of an agent
Claude Sonnet 5.5 costs half as much as Claude Opus 5.5, at $2 / $10 per 1M input / output tokens against $4 / $20, and Anthropic's own results put Sonnet 5.5 ahead on Terminal-Bench 4.0 and level on GDPval-AA, so Sonnet suits high-volume agent steps and Opus suits the steps that fail an eval on Sonnet.
Checked
In the Claude app, Opus or Sonnet is a choice from a model menu. For a team building on the API, it is a routing decision made per step of an agent: which tier each call should use, and what that does to the bill.
The 5.5 generation narrowed the gap. Sonnet 5.5 is half the price of Opus 5.5, both share the same 1M context and 128K output, and on Anthropic's own published results Sonnet 5.5 matches or leads Opus 5.5 on two agentic benchmarks.
The two models side by side
Claude Opus 5.5 (claude-opus-5-5) was released on 22 September 2026 at $4 / $20 per 1M tokens, down from $5 / $25 for Opus 5. Anthropic says it costs 40% less to run than Opus 5 on typical workloads, and a fast mode is available at $8 / $40. Anthropic's docs recommend starting with Opus 5.5 for most workloads.
Claude Sonnet 5.5 (claude-sonnet-5-5) was released on 28 September 2026 at $2 / $10, and Anthropic says it is up to 30% cheaper per task than Sonnet 5. Both models take 1M tokens of context with up to 128K output. Cache reads cost $0.20 per 1M on Opus 5.5 and $0.10 on Sonnet 5.5.
What Anthropic's own numbers show
These results come from Anthropic's own launch posts, so they compare like with like. On Terminal-Bench 4.0, Sonnet 5.5 scores 70.6% and Opus 5.5 scores 66.4%. On GDPval-AA v2.1, Sonnet 5.5 scores 1844 and Opus 5.5 scores 1846. Opus 5.5 testers also reported an unattended session of more than 18 hours.
Vendor benchmarks are a starting point. A narrow lead on a public benchmark says little about a specific agent's tasks, which is why the routing decision rests on an eval built from the team's own data.
Dates that affect migrations
Sonnet 4.5 retires on 30 November 2026
Anthropic names claude-sonnet-5-5 as the replacement. Agents still pinned to Sonnet 4.5 need to move and be re-evaluated before that date.
Earlier retirements
Opus 4 and Sonnet 4 retired on 15 June 2026, and Opus 4.1 on 5 August 2026.
Sampling parameters removed
API version 2026-10-21 removes temperature, top_p and top_k; requests that set them return a 400 error.
Haiku 5.5 below both
Released 7 October 2026 and positioned for subagents, browser use and high-volume work, Haiku 5.5 is the third option for steps that pass on neither price point.
How to decide
Start each step on the cheaper tier that might pass, run the eval, and move up only where it fails. Classification, extraction and routine tool calls often stay on Sonnet or Haiku; long planning and recovery from failed tool calls are where Opus is worth testing. The outcome is usually a mix, not one model.
Motiontic's engineers are vendor-neutral. They compare candidate models on the client's own data with an eval suite, route each step of the agent to the cheapest model that passes its eval, and leave the eval suite with the client, so the next model change is a measured decision rather than a guess.
Claude Opus 5.5 and Sonnet 5.5 compared. Prices in USD per 1M input / output tokens; benchmarks as reported by Anthropic; as of 31 October 2026
| Claude Opus 5.5 | Claude Sonnet 5.5 | |
|---|---|---|
| API ID | claude-opus-5-5 | claude-sonnet-5-5 |
| Released | 22 Sep 2026 | 28 Sep 2026 |
| Price | $4 / $20 (fast mode $8 / $40) | $2 / $10 |
| Cache read | $0.20 per 1M | $0.10 per 1M |
| Context / max output | 1M / 128K | 1M / 128K |
| Terminal-Bench 4.0 (Anthropic) | 66.4% | 70.6% |
| GDPval-AA v2.1 (Anthropic) | 1846 | 1844 |
| Against previous version | 40% less to run than Opus 5 on typical workloads | Up to 30% cheaper per task than Sonnet 5 |
Common questions
Is Claude Opus 5.5 better than Sonnet 5.5?
Not across the board. On Anthropic's own results, Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 against 66.4% for Opus 5.5, and the two are level on GDPval-AA v2.1 (1844 and 1846). Opus 5.5 costs twice as much. Which one fits a given agent step is settled by an eval on that step's real tasks.
How much do Claude Opus 5.5 and Sonnet 5.5 cost?
Claude Opus 5.5 costs $4 / $20 per 1M input / output tokens, with a fast mode at $8 / $40. Claude Sonnet 5.5 costs $2 / $10. Cache reads are $0.20 and $0.10 per 1M respectively.
Do Opus 5.5 and Sonnet 5.5 have the same context window?
Yes. Both take 1M tokens of context with up to 128K output, and the per-token price is the same for long prompts as for short ones.
What replaces Claude Sonnet 4.5?
Claude Sonnet 4.5 retires on 30 November 2026, and Anthropic names claude-sonnet-5-5 as the replacement. Agents pinned to Sonnet 4.5 should be moved and re-evaluated before that date.
When should an agent use Haiku 5.5 instead of Sonnet or Opus?
Claude Haiku 5.5 is positioned for subagents, browser use and high-volume work, and costs $0.10 / $0.50 per 1M tokens on prompts up to 100K tokens. It suits steps that pass their eval on the smallest tier, which keeps Sonnet and Opus for the steps that need them.
Related
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Claude Fable vs Opus
Anthropic's top two tiers compared on price, defaults and Anthropic's own results.
Claude vs ChatGPT
Plans, API prices, context limits and agent tooling for Anthropic's Claude and OpenAI's GPT models.
Eval
Repeatable, scored tests that tell a team whether an AI change made things better or worse.
LLM as a judge
A model grading model output against a rubric, checked against human labels.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.