Claude Opus 5.5 vs GPT-6 Sol: Capability, Cost, and the Right Workload

Claude Opus 5.5 and GPT-6 Sol arrived on the same day, but they solve different buying problems. Anthropic moved a lower-priced model into frontier territory. OpenAI pushed more of the GPT-6 generation into cheaper, faster tiers.

Watch Matt Wolfe’s First-Look Review

Source and credit: Opus 5.5 Is Crazy Good and GPT-6-Sol Launched Too, published by Matt Wolfe on 22 September 2026. The video is a rapid first look recorded on launch day; the official release pages below are the source of record for pricing, availability, and vendor-reported results.

Two Models Launched; Two Different Stories

Anthropic describes Opus 5.5 as the first model in its Claude 5.5 family. It says the model performs at the level of Fable 5.1 on most work, generates output more than 30% faster than Opus 5, and costs about 40% less than Opus 5 on typical workloads.

The interesting shift is product positioning. Opus used to imply a compromise below Anthropic’s highest Fable tier. Opus 5.5 reportedly beats Opus 5 and Fable 5.1 on nearly every benchmark in Anthropic’s launch table while costing far less than Fable 5.1’s $10 input and $50 output rates.

OpenAI’s Sol and Luna announcement makes a different promise. Astra remains the company’s best model. Sol is a highly capable, lower-cost alternative for coding, professional work, and agents; Luna is the fastest and least expensive GPT-6 option. The point is wider access to the GPT-6 generation, not a new model above Astra.

The Price Table Is Clear; the Invoice Is Not

These are standard list prices per million tokens on 22 September 2026. Anthropic also lists Opus 5.5 cache reads at $0.20 per million tokens, down 60% from Opus 5, and cache writes at $5. OpenAI says GPT-6 cached input receives a 90% discount and reports improved cache reuse for long agent conversations.

A cheaper token does not guarantee a cheaper completed task. A model that emits more tokens, retries a failed tool call, or needs more human correction can erase its list-price advantage. Matt’s review makes this point while looking at independent cost-per-task estimates: Opus 5.5’s token rate falls sharply, but high-effort runs can still consume a large amount of output.

What the Benchmarks Support, and What They Do Not

Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode 1.1, 57.8% on CursorBench 4.0, and 1846 Elo on GDPval-AA 2.1. Its table places the model ahead of Fable 5.1, Opus 5, and the listed GPT models on those tests. GPT-6 Astra remains ahead on AutomationBench and Terminal-Bench-Science in the same comparison.

OpenAI reports GPT-6 Sol at 33.2% on AutomationBench at xhigh effort for $0.27 per task. It also reports 56.4% on Agents’ Last Exam, 68.8% on DeepSWE 1.1 at max effort, and major improvements over GPT-5.6 Sol in factuality and coding efficiency. OpenAI explicitly says Astra remains its best model across the board.

These are useful signals, not a universal ranking. The vendors use different harnesses, effort settings, fallbacks, tools, and cost assumptions. Anthropic itself says benchmark margins at this level are becoming less reliable guides to real-world differences. Independent aggregators such as Artificial Analysis help by testing models under a shared framework, but their scores and prices are still snapshots that change as evaluations and serving behavior change.

The Creative Demos Are a Capability Probe

Matt highlights launch-day Opus 5.5 demos involving JavaScript animation, playable Game Boy and Snake-style experiences, first-person games, a flight simulator, a Mario Maker-like tool, and a clay-animation Blender render. They are impressive because they combine code generation, visual judgment, iteration, and execution across a longer task.

They do not prove that one prompt reliably produces production-ready games. Several examples came from early-access testers or Anthropic employees, and the public clips do not expose every prompt, retry, edit, tool, or elapsed minute. Treat them as evidence that the ceiling moved, then test the workflow you actually need.

For a creative team, the practical questions are more specific: can the model preserve a visual system across revisions, fix one component without breaking others, use the required toolchain, and produce an artifact another person can maintain?

Where the New Models Are Available

OpenAI’s launch page says Sol and Luna were not yet available in the standard Chat surface at launch and were rolling out gradually in Work and Codex. Availability can change after publication, so check the current model picker and official documentation before planning a customer workflow around a specific surface.

A Practical Routing Policy

The routing rule should live in your evaluation results, not in a permanent opinion about which lab is ahead. New models change quickly; your acceptance criteria can remain stable.

A Five-Task Adoption Test

1. Freeze five real tasks. Include one short task, one long task, one tool-heavy task, one ambiguous task, and one task where a subtle error would be costly.
2. Use the same context. Give each model the same files, tools, permissions, definition of done, and maximum retry count.
3. Run more than once. Three attempts per task reveal variance that a single polished demo hides.
4. Score the finished artifact. Track correctness, completeness, maintainability, visual quality where relevant, elapsed time, and reviewer intervention.
5. Calculate accepted-result cost. Include all tokens, retries, tool fees, and review time. Route each workload to the cheapest model that consistently clears the bar.

Video Chapters

Sources and Links

• Matt Wolfe: Opus 5.5 Is Crazy Good and GPT-6-Sol Launched Too
• Anthropic: Introducing Claude Opus 5.5
• Anthropic: Claude Opus 5.5 System Card
• OpenAI: Introducing GPT-6 Sol and Luna
• OpenAI Deployment Safety Hub: GPT-6 system-card material for Sol and Luna
• Artificial Analysis: independent model intelligence and cost comparisons