Claude Sonnet 5.5 Climbs to #2 on Artificial Analysis — Same $2/$10 List Price as Sonnet 5
Anthropic shipped Claude Sonnet 5.5 on September 28 at the same $2/$10 per-million-token rates as Sonnet 5. Artificial Analysis puts it at 56 on its Intelligence Index — two points behind Opus 5.5 — with migration breaks that will catch teams swapping only the model ID.
Anthropic released Claude Sonnet 5.5 on September 28, 2026 — six days after Claude Opus 5.5. The second model in the Claude 5.5 family keeps Sonnet 5’s list price and aims the upgrade at everyday coding, bug fixes, and polished documents rather than open-ended judgment work.
The rate card is unchanged: $2 per million input tokens, $10 per million output tokens, and $0.20 per million for cache reads, per Anthropic’s pricing docs. The API ID is claude-sonnet-5-5. It is live on the Claude Platform, Amazon Bedrock, Google Cloud, and Microsoft Azure.
Independent numbers landed the same day. Artificial Analysis scored Sonnet 5.5 at 56 on its Intelligence Index at max effort — second overall, two points behind Opus 5.5 and three ahead of GPT-6 Astra. Anthropic’s own launch table is more aggressive. Treat the vendor table as vendor-reported until your own harness agrees.
What Anthropic Actually Announced
In the launch post, Anthropic calls Sonnet 5.5 “a clear upgrade over Claude Sonnet 5” that “runs 30%+ faster” and “costs up to 30% less for most work.” Those speed and cost-per-task claims are Anthropic’s workload estimates, not a list-price cut.
Positioning is explicit. Opus 5.5 stays the model for complex work that needs sustained judgment. Sonnet 5.5 is for well-scoped tasks: fixing bugs, creating documents, slides, and spreadsheets, and fast iteration. Claude Haiku 5.5 is still promised “in the coming weeks.”
AWS confirmed the same day that Sonnet 5.5 is available on Amazon Bedrock via the Global CRIS inference profile, with example model ID global.anthropic.claude-sonnet-5-5, and on Claude Platform on AWS in North America. Zero data retention remains available, matching Opus 5.5 and Sonnet 5.
The system card lists a reliable knowledge cutoff of June 2026. Context window stays 1 million tokens with image and text input.
The Rate Card, Straight From Anthropic
Standard Claude API rates from Anthropic’s pricing page, in USD per 1 million tokens:
| Model | Input | Cache read | 5m cache write | Output |
|---|---|---|---|---|
Claude Sonnet 5.5 (claude-sonnet-5-5) |
$2 | $0.20 | $2.50 | $10 |
| Claude Sonnet 5 | $2 | $0.20 | $2.50 | $10 |
| Claude Opus 5.5 | $4 | $0.20 | $5 | $20 |
| GPT-6 Sol (OpenAI) | $2 | $0.20 | — | $10 |
One-hour cache writes for Sonnet 5.5 are $4 per million tokens. Batch API is half price: $1 / $5. That puts Sonnet 5.5 on the same list-price shelf as GPT-6 Sol — half of Opus 5.5 on input and output.
Anthropic’s “up to 30% less per task” line is not a cheaper token. It assumes fewer tokens per finished job. Artificial Analysis’s cost math, below, shows the opposite pattern at max effort. Run your own traffic before you budget against either story.
Vendor Benchmarks vs Artificial Analysis
Anthropic’s launch table is Anthropic’s own runs. Select scores from that post:
| Eval | Sonnet 5.5 | Sonnet 5 | Opus 5.5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 70.6% | 10.3% | 66.4% (Xhigh) |
| FrontierCode 1.1 (Main) | 46.2% Max / 52.1% Xhigh | 42.4% | 54.4% |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% |
| GDPval-AA v2.1 (Elo) | 1844 | 1449 | 1846 |
| OSWorld 2.1 (partial) | 80.1% | 57.0% | 81.8% |
Those are vendor-reported. Anthropic itself notes Opus 5.5 remains clearly stronger on complex, open-ended work even when scoreboards look close. A footnote also says Artificial Analysis ran GDPval-AA and AA-Briefcase on a pre-release build with a structured-outputs bug that has since been fixed — any effect, Anthropic says, would understate Sonnet 5.5.
Artificial Analysis’s independent write-up is the sharper check:
- Intelligence Index: 56 at max effort (#2), versus Opus 5.5 at the top and GPT-6 Astra / Claude Fable 5.1 tied lower at 53 in secondary coverage of the same leaderboard.
- Terminal-Bench 4.0: about 64% at max — ahead of Opus 5.5 (~60%) and GPT-6 Astra (~60%) in AA’s harness, but well below Anthropic’s own 70.6% figure.
- Token burn: ~193,000 output tokens per Intelligence Index task at max — the highest AA has measured, ~60% above Opus 5.5 or Sonnet 5 at max, and roughly 7× GPT-6 Astra.
- Cost per task: about $7.60 per Index task at max, ~50% higher than Sonnet 5 despite identical list prices.
- Pareto note: AA says Sonnet 5.5 sits off the intelligence-vs-cost frontier at several settings; high effort is its most competitive point, narrowly behind GPT-6 Sol for similar cost.
Same model, two scoreboards, different harnesses and effort settings. If you only read Anthropic’s Terminal-Bench jump from 10.3% to 70.6%, you will overstate what AA measured. Computer-use fans comparing OSWorld numbers should also read our Holo4 / OSWorld 2.0 coverage — those open agents trail Opus 5.5 by about 20 points on a related desktop-control bench.
Migration Breaks You Cannot Ignore
Anthropic’s migration guide is clear: swapping the model ID is not enough if you are coming from Sonnet 5.
thinking: {"type": "disabled"}returns a 400. On Sonnet 5.5 the lowest setting isthinking: {"type": "between_tools"}. Adaptive thinking is the default when the field is omitted.- Forced tool use is gone.
tool_choiceofanyor a named tool returns a 400. Useauto(plusstrict: trueon the Claude API where available). Bedrock does not offer strict structured outputs for Sonnet 5.5 yet — validate in your code. - Old computer-use tool rejected on Claude API and Google Cloud.
computer_20251124returns a 400; move tocomputer_toolset_20260801. Bedrock still acceptscomputer_20251124. - Thinking blocks are conversation-bound. For accounts created on or after August 31, 2026, editing earlier history after a thinking block can 400. Keep conversations append-only. Sonnet 5.5 does not read thinking from Opus 5 / Opus 5.5 / Fable / Mythos.
Default effort on the Claude Platform is high; Claude apps default to medium. Five levels exist: low, medium, high, xhigh, max. Levels are recalibrated versus Sonnet 5 — re-sweep before you lock a budget. Text between tool calls often arrives in thinking blocks now, so streaming UIs that only render text can go quiet mid-loop.
Safeguards Catch Up to the Capability Jump
Anthropic says Sonnet 5.5’s cybersecurity capabilities are comparable to Opus 5’s, so it is the first Sonnet to launch with cyber safeguards and fallbacks like those on the flagship line. Routine software development stays available; higher-risk cyber tasks fall back to Sonnet 5. Biology safeguards match Sonnet 5. Distillation / reasoning-extraction classifiers are new for the Sonnet tier.
On Anthropic’s automated behavioral audit (~1,850 scenarios), the company says Sonnet 5.5 improves on or matches Sonnet 5 on most alignment measures and is among the least likely of its models to probe container limits. That is Anthropic measuring Anthropic. The system card is the full package.
What It Means for Indian Developers
For Bangalore, Hyderabad, and Pune teams that already standardized on Sonnet 5 for volume work, Sonnet 5.5 is the obvious next ID — same $2 / $10 list price, same cache-read rate that dominates long agent loops. The catch is token volume at high effort: if AA’s ~193k output tokens per hard task shows up in your traces, your monthly bill can rise even when the rate card does not move.
A practical stack for cost-sensitive Indian startups right now:
- Sonnet 5.5 at medium / high for IDE agents, bug fixes, and document work where latency and polish matter.
- Opus 5.5 or GPT-6 Sol for multi-hour architecture and judgment-heavy coding.
- Open weights such as Qwen 3.8-27B for offline or air-gapped pipelines where API egress is the real cost.
Budget a day for migration. Forced tool choice and thinking: disabled will break Sonnet 5 integrations on day one. Teams on Cursor, Claude Code, or Bedrock should re-check computer-use tool versions and effort defaults before rolling the ID into production.
What’s Confirmed vs. Still Open
| Claim | Status |
|---|---|
| $2 / $10 API rates; $0.20 cache reads; Batch $1 / $5 | Confirmed on Anthropic pricing docs |
Model ID claude-sonnet-5-5; live on API, Bedrock, Google Cloud, Azure |
Confirmed by Anthropic and AWS |
| Migration breaks (disabled thinking, forced tools, computer toolset) | Confirmed in Anthropic’s migration guide |
| “30%+ faster” / “up to 30% less per task” | Anthropic workload claim — not a list-price identity |
| Terminal-Bench 70.6%, CursorBench 55.5%, OSWorld 80.1% | Vendor-reported from Anthropic’s launch post |
| Intelligence Index 56 (#2); ~193k output tokens/task at max | Independent — Artificial Analysis (pre-release build caveat) |
| Haiku 5.5 “in the coming weeks” | Anthropic promise — no date or price yet |
Frequently Asked Questions
How much does Claude Sonnet 5.5 cost?
On the Claude API: $2 per million input tokens and $10 per million output tokens — the same as Sonnet 5. Cache reads are $0.20; 5-minute cache writes are $2.50; 1-hour writes are $4. Batch is $1 / $5.
What is the Claude Sonnet 5.5 model ID?
claude-sonnet-5-5 on the Claude API. On Amazon Bedrock’s global profile, AWS shows global.anthropic.claude-sonnet-5-5. It is also on Google Cloud and Microsoft Azure.
Is Claude Sonnet 5.5 better than Claude Opus 5.5?
On Artificial Analysis’s composite index at max effort, Sonnet 5.5 is two points behind Opus 5.5. Anthropic still positions Opus for complex, open-ended work and Sonnet for well-scoped, faster, cheaper tasks. Scoreboards alone do not settle which model wins your workload.
Will my Claude Sonnet 5 code keep working?
Not if you disable thinking with the old flag, force tool choice, or still call computer_20251124 on the Claude API or Google Cloud. Update the model ID, switch to between_tools or adaptive thinking, use tool_choice: auto, and migrate computer use to the 2026 toolset. See Anthropic’s migration guide.
When does Claude Haiku 5.5 ship?
Anthropic only says “in the coming weeks.” No date or price has been published as of September 29, 2026.