Summary
Claude Sonnet 5.5, released by Anthropic on September 28, 2026, represents a dramatic generational leap over its predecessor, Sonnet 5. On the Terminal-Bench 4.0 agentic coding benchmark, Sonnet 5.5 scores 70.6% according to Anthropic (64% per Artificial Analysis independent evaluation), compared to Sonnet 5's 10.3% — a roughly sevenfold improvement [1]. The model achieves near-parity with Claude Opus 5.5 on knowledge-work benchmarks such as GDPval-AA v2.1 (Elo 1,844 vs. 1,846), while costing half as much per token ($2/$10 per 1M input/output tokens vs. Opus 5.5's $4/$20) [1]. Independent verification from Artificial Analysis shows smaller performance gaps between Sonnet 5.5 and its competitors than Anthropic's own vendor-published numbers suggest [1]. The model is not yet listed on LMSYS/LMArena or the Hugging Face Open LLM Leaderboard due to its extremely recent release, and no traditional academic benchmark scores (MMLU, GPQA, HumanEval) have been published for it yet [1].
Key Specifications
| Attribute | Value |
|---|---|
| Release Date | September 28, 2026 |
| API Model ID | claude-sonnet-5-5 |
| Context Window | 1,000,000 tokens |
| Max Output | 128,000 tokens |
| Pricing (input) | $2.00 per 1M tokens |
| Pricing (output) | $10.00 per 1M tokens |
| Adaptive Thinking | On by default (lowest: between_tools) |
| Effort Levels | Low, Medium, High, XHigh, Max |
| Output Generation Speed | 30%+ faster than Sonnet 5 |
| Cost per Task (typical effort) | Up to 30% lower than Sonnet 5 |
| Cybersecurity Safeguards | Yes (first Sonnet model with these) |
| Anti-Distillation Protection | Yes |
The adaptive thinking system dynamically adjusts reasoning depth based on task complexity, with five effort levels giving developers fine-grained control over the cost-performance tradeoff. At the lowest effort setting (between_tools), the model prioritizes speed; at Max effort, it allocates the most compute to each token [1].
2. Benchmark Results
2.1 Terminal-Bench 4.0 (Agentic Coding)
Terminal-Bench 4.0 tests multi-hour autonomous terminal work — the most demanding agentic coding benchmark currently in use. Sonnet 5.5 shows the largest relative improvement over its predecessor of any benchmark in this dataset.
| Model | Score (Anthropic) | Score (AA Independent) | Delta vs. Sonnet 5 |
|---|---|---|---|
| Claude Sonnet 5.5 | 70.6% | 64.0% | +60.3 pp / +50.0 pp |
| Claude Sonnet 5 | 10.3% | 14.0% | — |
| Claude Opus 5.5 | 66.4% | 60.0% | — |
| GPT-6 Sol | — | 44.0% | — |
The gap between Anthropic's vendor-published score (70.6%) and the Artificial Analysis independent evaluation (64.0%) is 6.6 percentage points, a pattern seen across multiple benchmarks and discussed further in Section 4 [1].
2.2 SWE-bench Pro
SWE-bench Pro evaluates models on 1,865 tasks across 41 private and copyleft code repositories, making it significantly harder than SWE-bench Verified (where top models score ~80%). The Pro variant better reflects real-world autonomy for coding agents.
| Model | Score |
|---|---|
| Claude Sonnet 5.5 | 81.3% |
| Claude Sonnet 5 | 63.2% |
| Claude Opus 5.5 | 89.9% |
Sonnet 5.5 closes much of the gap to Opus 5.5 compared to Sonnet 5, but Opus 5.5 still leads by 8.6 percentage points [1].
2.3 GDPval-AA v2.1 (Elo Rating)
GDPval-AA v2.1 measures model performance on tasks that simulate real-world economic decision-making across 44 occupations, evaluated by human experts using the Elo rating system.
| Model | Elo Rating |
|---|---|
| Claude Opus 5.5 | 1,846 |
| Claude Sonnet 5.5 | 1,844 |
| GPT-6 Sol | 1,487 |
| Claude Sonnet 5 | 1,449 |
Sonnet 5.5 achieves near-parity with Opus 5.5 on knowledge work, trailing by only 2 Elo points — a negligible difference [1].
2.4 AA-Briefcase v1.1 (Elo Rating)
AA-Briefcase v1.1 is an independent evaluation of model performance on real-world business and professional tasks, also using the Elo system.
| Model | Elo Rating |
|---|---|
| Claude Opus 5.5 | 1,822 |
| Claude Sonnet 5.5 | 1,811 |
| GPT-6 Sol | 1,483 |
| Claude Sonnet 5 | 1,359 |
Again, Sonnet 5.5 trails Opus 5.5 by only 11 Elo points while outperforming GPT-6 Sol by 328 points [1].
2.5 Artificial Analysis Intelligence Index v4.3.2
The AA Intelligence Index is an independent composite score that aggregates performance across multiple benchmarks to produce a single intelligence ranking.
| Rank | Model | Score |
|---|---|---|
| 1 | Claude Opus 5.5 | 58 |
| 2 | Claude Sonnet 5.5 | 56 |
| 3 | GPT-6 Astra | 53 |
| 4 | Claude Sonnet 5 | 38 |
Sonnet 5.5 ranks second on the AA Intelligence Index, 2 points behind Opus 5.5 and 3 points ahead of GPT-6 Astra [1].
2.6 BenchAlign Composite (benchlm.ai, v5.7)
BenchAlign is a composite ranking from benchlm.ai that aggregates scores across multiple benchmarks into a single score.
| Rank | Model | BenchAlign Score |
|---|---|---|
| 1 | GPT-6 Astra | 88.5 |
| 2 | Claude Opus 5.5 | 87.07 |
| 3 | Claude Fable 5.1 | 82.96 |
| 4 | GPT-6 Sol | 81.08 |
| 5 | Claude Sonnet 5.5 | 80.49 (estimated) |
| 6 | Claude Opus 5 | 79.64 |
The Sonnet 5.5 BenchAlign score is estimated rather than directly sourced from benchlm.ai, as the model was released after the v5.7 report was finalized [1].
2.7 LMSYS/LMArena Chatbot Arena
LMArena is the largest community-driven LLM evaluation platform, using blind pairwise comparisons to produce Elo ratings. Sonnet 5.5 is not yet listed due to its release on September 28, 2026 — the day before this report was compiled.
| Model | LMArena Text Elo |
|---|---|
| Claude Mythos 5 | 1,531 |
| Claude Fable 5 | 1,525 |
| Claude Opus 5 | 1,522 |
| GPT-5.6 Sol | 1,514 |
| Claude Sonnet 5 | 1,479 |
Sonnet 5.5's absence from LMArena is expected given its one-day age. Its predecessor, Sonnet 5, sits at #18 with 1,479 Elo [1].
3. Additional Notable Benchmarks
Beyond the headline benchmarks, Sonnet 5.5 shows strong performance across a range of specialized evaluations.
| Benchmark | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|
| OSWorld 2.1 (partial) | 80.1% | 57.0% | 81.8% | — |
| CursorBench 4.0 | 55.5% | 34.1% | 57.8% | — |
| FrontierCode 1.1 (XHigh) | 52.1% | — | — | — |
| FrontierCode 1.1 (Max) | 46.2% | — | — | — |
| AutomationBench | 71% | 37% | 70% | — |
| HLE (with tools) | 64.5% | 54.9% | 67.7% | — |
| Chartography (no tools) | 61.6% | 15.6% | 64.4% | — |
| HealthBench Professional | 69.2% | 57.8% | 65.6% | — |
| SWE-Bench Multilingual | 90.3% | 78.3% | 93.9% | — |
| ProgramBench (long context) | 79.7% | 77.3% | 91.2% | — |
Several patterns emerge from this broader set of results. Sonnet 5.5 shows the largest relative gains over Sonnet 5 on benchmarks that involve tool use and agentic behavior (Chartography: 61.6% vs. 15.6%; AutomationBench: 71% vs. 37%). On benchmarks that test sustained reasoning or long-context coding (ProgramBench, HLE), the gap to Opus 5.5 remains wider. The model also shows particular strength in multilingual software engineering (SWE-Bench Multilingual: 90.3%) [1].
4. Competitor Models
4.1 GPT-6 Astra (OpenAI)
GPT-6 Astra is OpenAI's current flagship, released September 3, 2026. It leads on the BenchAlign composite (88.5) and the AA Intelligence Index (53 at max effort), and holds the top GPQA Diamond score (96.1% at XHigh effort) [1][2].
| Attribute | GPT-6 Astra |
|---|---|
| Pricing (input/output per 1M tokens) | $10 / $50 |
| Context Window | ~1,050,000 tokens |
| AA Intelligence Index (max) | 53 |
| BenchAlign (v5.7) | 88.5 |
| GPQA Diamond (XHigh) | 96.1% |
| Terminal-Bench 4.0 (AA) | 57.9% |
| LMArena Text Elo | 1,479.8 |
GPT-6 Astra's per-token cost is 5× higher than Sonnet 5.5 on input and 5× higher on output. Its strengths lie in hard science reasoning (GPQA Diamond) and the BenchAlign composite, while Sonnet 5.5 leads on agentic coding (Terminal-Bench 4.0) and costs significantly less [1][2].
4.2 GPT-6 Sol (OpenAI)
GPT-6 Sol is OpenAI's more affordable tier, priced identically to Sonnet 5.5 at $2/$10 per 1M tokens. It serves as a direct cost-competitor.
| Attribute | GPT-6 Sol |
|---|---|
| Pricing (input/output per 1M tokens) | $2 / $10 |
| AA Intelligence Index | 48 |
| BenchAlign (v5.7) | 81.08 |
| GDPval-AA v2.1 (Elo) | 1,487 |
| Terminal-Bench 4.0 (AA) | 44.0% |
| LMArena Text Elo | ~1,483 (estimated) |
At equal pricing, Sonnet 5.5 outperforms GPT-6 Sol on nearly every benchmark where both have published scores: GDPval-AA Elo (1,844 vs. 1,487), Terminal-Bench 4.0 (64.0% vs. 44.0% per AA), and AA-Briefcase Elo (1,811 vs. 1,483) [1].
4.3 Claude Opus 5.5 (Anthropic)
Opus 5.5 is Anthropic's flagship, released September 22, 2026 — six days before Sonnet 5.5. It leads the AA Intelligence Index (58) and the BenchAlign composite (87.07) [1][2].
| Attribute | Opus 5.5 | Sonnet 5.5 |
|---|---|---|
| Pricing (input/output per 1M tokens) | $4 / $20 | $2 / $10 |
| AA Intelligence Index (max) | 58 | 56 |
| BenchAlign (v5.7) | 87.07 | 80.49 (est.) |
| SWE-bench Pro | 89.9% | 81.3% |
| Terminal-Bench 4.0 (Anthropic) | 66.4% | 70.6% |
| Terminal-Bench 4.0 (AA independent) | 60.0% | 64.0% |
| GDPval-AA v2.1 (Elo) | 1,846 | 1,844 |
| Cost per task (Max effort) | $5.98 | $7.60 |
The cost comparison is nuanced. At typical effort levels, Sonnet 5.5 is significantly cheaper and faster than Opus 5.5, with 30%+ faster output generation and up to 30% lower cost per task [1]. However, at Max effort, the cost per task for Sonnet 5.5 ($7.60) actually exceeds that of Opus 5.5 ($5.98), because Sonnet 5.5's lower per-token price is offset by its higher token consumption at the highest effort level [1].
4.4 Claude Fable 5.1 (Anthropic)
Fable 5.1 is Anthropic's long-horizon agent model, released September 1, 2026. It leads on the AA Intelligence Index at max effort among the Fable lineage (53.4) and on the LMArena text arena (1,505.7 Elo for Fable 5 high) [2].
| Attribute | Fable 5.1 | Sonnet 5.5 |
|---|---|---|
| Pricing (input/output per 1M tokens) | $10 / $50 | $2 / $10 |
| AA Intelligence Index (max) | 53.4 | 56 |
| BenchAlign (v5.7) | 82.96 | 80.49 (est.) |
| SWE-bench Pro | 80.0% | 81.3% |
| Terminal-Bench 4.0 | 55.8% | 70.6% (Anthropic) |
Sonnet 5.5 outperforms Fable 5.1 on Terminal-Bench 4.0 and SWE-bench Pro while costing 5× less per input token and 5× less per output token [1].
5. Key Insights
5.1 Sonnet 5.5 Dramatically Outperforms Sonnet 5 on Agentic Coding
The most striking finding is the gap between Sonnet 5.5 and Sonnet 5 on Terminal-Bench 4.0: 70.6% vs. 10.3% per Anthropic's own evaluation. Even the independent Artificial Analysis score of 64.0% represents a roughly 4.6× improvement over Sonnet 5's 14.0% [1]. This is not a marginal upgrade — it is a new model class for agentic terminal work.
5.2 Near-Parity with Opus 5.5 on Knowledge Work
On GDPval-AA v2.1, Sonnet 5.5 achieves an Elo of 1,844 against Opus 5.5's 1,846 — a difference of 2 points that is statistically negligible [1]. On AA-Briefcase v1.1, the gap widens slightly to 11 points (1,811 vs. 1,822). This suggests that Sonnet 5.5 has been tuned to close the knowledge-work gap with Opus 5.5 while maintaining its cost advantage.
5.3 Significant Cost Advantage Over Opus 5.5 at Typical Effort Levels
At $2/$10 per 1M tokens vs. Opus 5.5's $4/$20, Sonnet 5.5 is 2× cheaper on both input and output. Combined with 30%+ faster output generation, the cost per task at typical (low/medium) effort levels is substantially lower [1]. This makes Sonnet 5.5 the more economical choice for workloads that do not require Opus 5.5's peak reasoning depth.
5.4 Not Yet on LMSYS/LMArena or Hugging Face
Sonnet 5.5 was released on September 28, 2026 — one day before this report. It is not yet listed on LMSYS/LMArena (the largest community evaluation platform) or the Hugging Face Open LLM Leaderboard (which excludes closed-source models). This means there is no independent community validation of its chat or general capabilities at the time of writing [1].
5.5 Independent Verification Shows Smaller Gaps
Artificial Analysis independent scores consistently show smaller gaps between Sonnet 5.5 and its competitors than Anthropic's vendor-published numbers. For example, on Terminal-Bench 4.0, Anthropic reports 70.6% while AA reports 64.0% — a 6.6-point difference. On Opus 5.5, Anthropic reports 66.4% while AA reports 60.0% — a 6.4-point difference [1]. This pattern suggests that vendor-published benchmarks may use more favorable evaluation setups, and independent verification should be weighted when making production decisions.
5.6 Cost Per Task at Max Effort Is Higher Than Opus 5.5
At Max effort, Sonnet 5.5's cost per task ($7.60) exceeds Opus 5.5's ($5.98). This is counterintuitive given Sonnet 5.5's lower per-token price, but it reflects higher token consumption at the highest effort level. Teams that routinely use Max effort should benchmark both models for their specific workload before concluding that Sonnet 5.5 is cheaper [1].
5.7 At Low and Medium Effort, Sonnet 5.5 Is Significantly Cheaper and Faster
The 30%+ faster output generation and up to 30% lower cost per task at typical effort levels make Sonnet 5.5 the clear choice for high-throughput, cost-sensitive workloads that do not need the deepest reasoning [1].
5.8 First Sonnet Model with Cybersecurity Safeguards
Sonnet 5.5 is the first model in the Sonnet family to include cybersecurity safeguards and anti-distillation protection. This is a notable addition for teams deploying the model in production environments where security posture matters [1].
6. Limitations
Several limitations should be considered when interpreting these results.
No traditional academic benchmark scores. MMLU, GPQA Diamond, HumanEval, and other well-known academic benchmarks have not been published for Sonnet 5.5. This means its performance on standardized knowledge and coding tests cannot be directly compared to older models on those specific evaluations [1].
Most scores are vendor-published or from a single independent source. The majority of Sonnet 5.5 scores come from Anthropic's own evaluation harness. Artificial Analysis provides independent verification for a subset of benchmarks, but the coverage is limited [1].
LMSYS/LMArena data not yet available. The model was released one day before this report was compiled. Community-driven evaluation on LMArena will take weeks to accumulate meaningful vote counts [1].
Hugging Face Open LLM Leaderboard excludes closed-source models. Sonnet 5.5 cannot appear on this leaderboard regardless of its performance [1].
BenchAlign score for Sonnet 5.5 is estimated. The v5.7 BenchAlign report was finalized before Sonnet 5.5's release. The 80.49 score is an estimate based on the model's performance on individual benchmarks that are included in the composite [1].
Benchmark comparability across effort levels. Different benchmarks use different effort settings and evaluation harnesses. Direct comparison between scores should account for the effort level at which each score was achieved. The AA Intelligence Index and BenchAlign composite attempt to normalize across these differences, but no normalization is perfect [1].
Terminal-Bench 4.0 vendor vs. independent gap. The 6.6-point gap between Anthropic's vendor score (70.6%) and AA's independent score (64.0%) for Sonnet 5.5 is consistent with a similar gap for Opus 5.5 (66.4% vs. 60.0%). This suggests a systematic upward bias in vendor-published Terminal-Bench scores, but the magnitude of the bias is not yet fully characterized [1].
7. Recommendations
Based on the available evidence, the following recommendations can be made for teams evaluating Sonnet 5.5:
For agentic coding workloads, Sonnet 5.5 is the strongest value proposition in the frontier class at $2/$10 per 1M tokens. Its Terminal-Bench 4.0 score of 64.0% (AA independent) is 46% higher than GPT-6 Sol's 44.0% at the same price point [1].
For knowledge work and professional tasks, Sonnet 5.5 offers near-parity with Opus 5.5 at half the cost. The 2-point Elo gap on GDPval-AA (1,844 vs. 1,846) is negligible for most production use cases [1].
For Max-effort workloads, benchmark both Sonnet 5.5 and Opus 5.5 for your specific task. Sonnet 5.5's cost per task at Max effort ($7.60) exceeds Opus 5.5's ($5.98), so the cost advantage reverses at the highest effort level [1].
Wait for LMArena and community benchmarks before making a final decision if your workflow depends on open, community-validated evaluations. The model's LMArena ranking will emerge over the coming weeks [1].
Request independent evaluation for your specific use case. The gap between vendor-published and independent scores (6–7 points on Terminal-Bench 4.0) suggests that internal benchmarking on your own tasks is essential [1].