Claude Sonnet 5.5 vs the AI Field: Opus 5.5, GPT-6 Astra, Gemini, Grok & the Open-Weight Tier

01 The Star: Claude Sonnet 5.5

Claude Sonnet 5.5 shipped on September 28, 2026 — six days after Opus 5.5 — as the second model in Anthropic's 5.5 family. It is a 1.0M-token context, 128K-output model priced at $2 per million input tokens, $10 per million output, and $0.20 per million cached-input tokens.

Anthropic's headline is that it runs 30% or more faster than Sonnet 5 and cuts cost per task by up to 30%. The cheaper bills come from doing the job in fewer steps and batching tool calls more aggressively — not from a lower sticker price, which Sonnet 5 already shared.

• FrontierCode 1.1 — 52.1 (Opus 5.5: 54.4, GPT-6 Sol: 49.3)
• CursorBench 4.0 — 55.5, within two points of Opus 5.5's 57.8
• GDPval-AA v2.1 — 1,844 Elo, two points behind Opus 5.5's 1,846 and 357 ahead of GPT-6 Sol
• System card: SWE-bench Pro 81.3, DeepSWE v1.1 71.0, OSWorld 2.1 at 43.5 strict / 80.1 partial

Default effort is high on the API and platform, medium inside the Claude apps and Claude Code. Adaptive thinking is on by default.

02 The Anthropic Family: Opus 5, Opus 5.5, Fable 5.1

Opus 5 (July 24, 2026) is the older flagship: $5/$25, 1M context, 128K output, default effort high. It is still in the API as the prior generation.

Opus 5.5 (September 22, 2026) replaced it as the default flagship — a 20% price cut to $4/$20, 1M context, 128K output, default effort medium, and adaptive thinking that can no longer be turned off. On Anthropic's own table it tops SWE-bench Pro at 89.9, Terminal-Bench 4.0 at 66.4, FrontierCode 1.1 at 54.4, CursorBench 4.0 at 57.8 and GDPval-AA at 1,846 Elo. Cache reads fall to 5% of input price ($0.20, the same as Sonnet 5), which is where the real agent-loop savings land.

Above it sits Fable 5.1, the $10/$50 long-horizon reasoning tier. Fable 5.1 and Mythos 5.1 are the same model with different safeguards — Fable is generally available, Mythos is trusted-access only. On Anthropic's published rows Opus 5.5 now beats Fable 5.1 on most of them, which is why the docs now say to start with Opus 5.5 and reserve Fable for the hardest, longest-running reasoning.

03 The OpenAI GPT-6 Line: Astra, Sol, Luna

"GPT Astra" is a family name, not a model. OpenAI's GPT-6 generation runs on three tiers, all released in September 2026, all with a 1.05M-token context and a 128K output cap.

source: openai gpt-6 model reference & api changelog, sep 2026 · gpt-6.1 sol shipped sep 29 at the same headline price

Astra dropped the "off" reasoning setting — every call spends thinking tokens. On the Artificial Analysis Intelligence Index, Opus 5.5 leads at 58, GPT-6 Astra scores 53, and Sol sits at 57. Sol's GDPval-AA is 1,487 and FrontierCode 49.3: strong, but a clear step below the flagships. Sol is priced exactly half of GPT-5.6 Sol and scores about the same — it is the upgrade path for anyone who does not need to pay for Astra.

04 The Rest of the Closed Frontier

Gemini 3.1 Pro (Google, February 19, 2026) is the current flagship here: a 1,048,576-token context with 65,536-token output, $2/$12 under 200K context, and native audio and video input. It scores 80.6 on SWE-bench Verified, 54.2 on SWE-bench Pro, 68.5 on Terminal-Bench 2.0 and 94.3 on GPQA Diamond — 48 on the AA Intelligence Index. It is the model to reach for when the context or the modality is the bottleneck.

Grok 4.6 (xAI, August 12, 2026) is the coding-and-agents pick: 500K context, $2/$6 under 200K, and 65.8 tok/s. It leads the AA Intelligence Index at 61, posts 57.5 on APEX-Agents, 1,753 on GDPval-AA and 69.9 on CursorBench v3.2. Grok and Gemini are priced to compete, and Grok costs less on output-heavy workloads.

05 The Open-Weight Tier

The story of 2026 is open weights. Trillion-parameter MoE models now ship under MIT or Apache for a fraction of the closed price.

• Qwen3.8-Max (Alibaba, API Aug 3, 2026) — 2.4T-parameter MoE, 95B active per token, ~1M context, 131,072-token output, $2/$6. Open weights were announced as "coming" in August; no confirmed Hugging Face drop appears in our sources as of late September.
• GLM-5.2 (Z.ai, June 2026) — 753B MoE under MIT, 1M context, $0.75/$2.40. It scores 62.1 on SWE-bench Pro (ahead of GPT-5.5) and jumped to 81.0 on Terminal-Bench from 62.0 the generation before.
• DeepSeek V4-Pro — 1.6T MoE, 49B active, MIT-licensed, ~$0.30/$1.20. The cost-efficiency reference point for self-hosting.
• Kimi K3 (Moonshot, open weights July 27, 2026) — 2.8T model, 896 experts with 16 active, 1M context, $3/$15. Currently rated around 57 on the AA index — the top open-weight result.

06 The Full Comparison

Thirteen models, one table. Specs and list pricing from each vendor; the "notes" column carries one verified anchor benchmark per model where one exists.

source: vendor model reference pages & system cards, sep 2026 · benchmark anchors from the row's own source, not cross-harness

07 The Bottom Line

• Everyday coding, agents and documents: Sonnet 5.5 is the default — Opus-class results on Terminal-Bench and GDPval at $2/$10.
• When a task needs the ceiling: Opus 5.5. Its 89.9 SWE-bench Pro and always-on thinking are the reason it exists.
• Hardest, longest-running reasoning: Fable 5.1 — and it is priced like a $10/$50 model.
• OpenAI workloads: Astra is the ceiling, Sol is the value pick, Luna is for high-volume, low-stakes jobs.
• Cross-vendor: Gemini 3.1 Pro wins on context and multimodal input; Grok 4.6 on agentic coding and the AA index.
• Self-hosting and budget: the open tier — GLM-5.2, DeepSeek V4-Pro, Kimi K3 — is the cheaper frontier, and Qwen3.8-Max is the one to watch once its weights land.