Claude Sonnet 5.5, released by Anthropic on September 28, is about two points behind Opus 5.5 on most of its benchmarks and costs half as much. That makes model selection in Claude Code a cost decision, not a capability one. But half the price is not half the cost, and in that difference lies the real decision.
A bit of context before the numbers: Claude Code doesn’t use Sonnet by default. On Pro, Max, Team, and Enterprise plans, and on the Anthropic API, the default model is Opus 5.5. So the question most teams are asking themselves this week isn’t “what model do I start with?”, but “is it worth dropping down from Opus, and for what work?”.
What is Claude Sonnet 5.5?
Claude Sonnet 5.5 is the second model in Anthropic’s Claude 5.5 family, designed as the faster and cheaper complement to Opus 5.5. Anthropic says its strengths are “well-scoped everyday tasks, bug fixes, and polished document creation”, and that versus Sonnet 5 it “runs 30%+ faster, and costs up to 30% less for most work”.
It became available to everyone on the day of the announcement: on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. The model ID is claude-sonnet-5-5, with a 1M token context window, maximum output of 128K, knowledge through June 2026, and no retirement before September 28, 2027.
Anthropic announced a third model, Haiku 5.5, for “the coming weeks”. At the time of publishing this note, it’s still not available.
Claude Sonnet vs Opus: How close is Sonnet 5.5 to Opus 5.5?
Very close according to Anthropic’s own table, and even ahead on a programming benchmark. All figures are reported by Anthropic; there’s no independent replication yet.
| Benchmark | Sonnet 5.5 | Opus 5.5 | Sonnet 5 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 70.6% | 66.4%¹ | 10.3% |
| CursorBench 4.0 | 55.5% | 57.8% | 34.1% |
| FrontierCode 1.1 (Main) | 52.1% (Xhigh) | 54.4% | 42.4% |
| GDPval-AA v2.1² | 1844 | 1846 | 1449 |
| OSWorld 2.1 (computer use) | 80.1% | 81.8% | 57.0% |
¹ Opus 5.5 reported at Xhigh effort, its best score. ² GDPval-AA was run by Artificial Analysis on a pre-release deployment that, according to Anthropic, had a structured outputs bug that has since been fixed.
The jump in Terminal-Bench, from 10.3% to 70.6%, is the headline, and it’s the figure that makes Sonnet 5.5 a serious option within a terminal agent. Sonnet 5 wasn’t.
But Anthropic is unusually direct about the limitation. The announcement says the benchmarks “capture only one facet” and that “Opus 5.5 remains clearly stronger at complex, open-ended work requiring sustained judgment”. The system card goes further: Sonnet 5.5 “usually does not reach the level of Opus 5.5”. I read the table as “Sonnet 5.5 can now do almost everything Opus does in well-defined tasks”, not as “Sonnet 5.5 equals Opus”.
What’s the price of Claude Sonnet 5.5 versus Opus?
Half, on almost every line of the price list:
| Per 1M tokens | Sonnet 5.5 | Opus 5.5 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Cache write | $2.50 | $5 |
| Cache read | $0.20 | $0.20 |
The last row is the one that matters. Cache reads cost the same on both models. And an agentic coding session is, above all, cache reads: each turn resends the conversation, the repository context, and tool results, and almost everything comes back from cache.
An illustrative turn (my arithmetic, not a measurement): 100K tokens read from cache, 5K new input, and 2K output.
- Opus 5.5: $0.020 cache + $0.020 input + $0.040 output = $0.080
- Sonnet 5.5: $0.020 cache + $0.010 input + $0.020 output = $0.050
That’s about 37% cheaper, not 50%. The longer the session and the larger the cached context, the closer the two become. The savings that survive come from output tokens, and that’s why Sonnet 5.5’s token efficiency weighs more than its list price.
Anthropic’s cost-per-task graphs point in the same direction. In their words, Sonnet 5.5 “complements Opus 5.5 best when running at lower effort settings, where it costs less per task. At higher settings, it can perform comparably at a similar cost”. In other words: if you run Sonnet 5.5 at Xhigh to match Opus, you’re largely paying the Opus bill again.
What’s the best Claude model for coding: Sonnet 5.5 or Opus 5.5?
My read, for a team already working with Claude Code:
- Keep Opus 5.5 as default for ambiguous work: architecture decisions, debugging with no clear cause, changes across multiple repositories, anything where you’d want a senior engineer to push back on you. That’s exactly the category that, according to Anthropic, Opus still dominates.
- Move well-scoped work to Sonnet 5.5 at medium effort: implementing a specified feature, fixing a bug with a known reproduction, writing tests, refactors with a clear goal. That’s where you notice the speed and the cost difference is bigger.
- Consider
opusplan, a Claude Code mode that uses Opus in plan mode and switches to Sonnet for execution. It’s the cleanest way to apply that split without thinking about it on every task. - For API workloads outside Claude Code (review bots, batch pipelines, CI agents), Sonnet 5.5 becomes the obvious first candidate. CodeRabbit, cited by Anthropic, says it plans to “move simple and moderate reviews over now”. It’s customer testimony published by the vendor itself, so weigh it as such, but it’s the pattern I’d expect.
The strategic point for CTOs is that the default model decision changed categories. In the previous generation, Sonnet was the economical option where you accepted quality loss. In this one, the gap on scoped work is so small that model selection becomes a routing policy, and routing policies go into team conventions, not individual developer habits.If your team also compares agents, the same reasoning applies across providers: we saw it in Claude Code vs Codex and in the launch of Gemini 3.7 Flash.
How to use Sonnet 5.5 in Claude Code?
Sonnet 5.5 requires Claude Code v2.1.284 or higher; first run claude update. Then:
- Within a session:
/model sonnet - At startup:
claude --model sonnet - Effort:
/effort(in Claude Code, Sonnet 5.5 usesmediumby default, same as Opus 5.5)
Note: /model saves your choice as the default for new sessions. To change only in the current session, open the selector with /model and press s on the model row.
One detail for enterprise readers: the sonnet alias only points to Sonnet 5.5 on the Anthropic API. At the time of publishing this note, it points to Sonnet 4.5 on Amazon Bedrock and Google Cloud, and to Sonnet 4.6 on Claude Platform on AWS. On those providers, select the model explicitly or set it with ANTHROPIC_DEFAULT_SONNET_MODEL.
Does Sonnet 5.5 Refuse Cybersecurity Tasks?
More so than Sonnet 5. Sonnet 5.5 is the first Sonnet to ship with cybersecurity safeguards similar to Opus 5.5, and Anthropic’s system card states it plainly: “users should expect increased refusals with Sonnet 5.5, even on benign cybersecurity-related tasks”. In Claude Code, requests flagged for cybersecurity automatically fall back to Sonnet 5; those flagged for biology end in refusal, with no fallback model. If your team does security research, fuzzing, or binary analysis, test it before switching.
What Changes for a Technical Team?
Sonnet 5.5 doesn’t replace Opus 5.5. What it does is make it worth intentionally answering the question “which model?”. On scoped tasks with medium effort it’s faster and considerably cheaper. On long, cache-heavy sessions, the savings shrink, and on open-ended work that demands judgment, Anthropic itself says Opus stays ahead. Treat it as a routing decision, document the rule, and revisit it when Haiku 5.5 arrives.