GPT-6 Sol and Claude Opus 5.5 Released: Coding Gains, Price Drops, and 3D Showcases

Simultaneous Launch of GPT-6 Sol and Claude Opus 5.5

OpenAI and Anthropic released new models on the same day.

GPT-6 Sol: Astra-Level Capabilities at Lower Cost

OpenAI positions Sol below Astra. Astra handles the hardest tasks; Sol brings this generation's advances in coding, computer operation, and factual accuracy to a price point suited for frequent calls.

Pricing

Sol's input and output token prices are both half of the previous generation. OpenAI also introduced a cheaper GPT-6 Luna, but this article focuses on Sol.

Coding Benchmarks

According to OpenAI's release notes [1], Sol shows clear improvement on FrontierCode 1.1 Main, which evaluates whether code can be merged into real repositories, checking test quality, code style, and adherence to project conventions.

On DeepSWE 1.1, a benchmark for complex long-horizon software engineering tasks, Sol scores 68.8% at the max tier. The author notes these tests reflect real-world agent workflows: writing features, adding tests, checking change scope, and respecting existing code constraints.

Factual Accuracy

OpenAI evaluated Sol on conversations previously flagged for factual errors. Sol's error count is roughly half that of the previous model. However, the test uses a curated error-prone sample; everyday conversation improvement remains to be seen.

Prompt Caching Improvements

GPT-6 increases prompt cache hit rates. Cache reads receive a 90% discount. Changing reasoning effort or adding/removing tools mid-conversation still reuses prior context. For example, an agent that first reads a repository, then edits code and runs tests, can reuse the project rules and history across subsequent requests, reducing repeated read costs. Savings depend on cache hits, output volume, and rework frequency.

A new official caching dashboard [2] shows hit-rate trends over time and breaks down cached reads, writes, and uncached inputs.

Availability

Sol and Luna are available in ChatGPT Work, Codex, and API (model ID: gpt-6-sol), but not yet in regular Chat.

Claude Opus 5.5: Faster Long Tasks, Clearer Responses

Anthropic positions Opus 5.5 [3] as matching Fable 5.1 on most tasks at lower cost than Opus 5.

Pricing and Speed

Standard API input/output prices dropped 20%. The claimed "40% cheaper typical tasks" includes reduced token consumption under default settings. Output generation speed is over 30% faster than Opus 5.

Large-Scale Code Migration Case Study

Anthropic tasked Opus 5.5 and Fable 5.1 with migrating HAProxy from C to Rust. Both passed nearly all original regression tests. Opus 5.5 completed the task in 9.5 hours vs. 12 hours for Fable 5.1, at 51% lower cost. The author highlights that using the original project's regression suite as acceptance criteria is valuable: migrating mature software requires preserving existing behavior, not just generating lines of code. However, passing "almost all" tests does not yet guarantee production readiness.

Terminal-Bench 4.0 Results

Opus 5.5 scores 66.4% at xhigh tier; Opus 5 scored 52.3% at max. These are Anthropic's self-reported numbers, not a head-to-head comparison with GPT-6 Sol in the same environment.

Response Clarity

Opus 5.5 places key information first, reduces jargon, and follows user-specified writing rules better. For long tasks where the model edits many files, clearer summaries reduce back-and-forth clarification.

Subscription Changes

Pro, Max, Team, and per-seat Enterprise plans receive increased five-hour usage limits plus a one-time rollover credit reset. API model ID: claude-opus-5-5.

Early 3D Demos on X

Addy Osmani's Three.js Cycling Pelican

Addy Osmani shared a Three.js demo [4] (377k views at screenshot time) of a pelican cycling through a streetscape with buildings, trees, parked cars, speed and progress indicators. A second variant shows a helmeted, sunglasses-wearing pelican with a red scarf cycling along a beach at sunset, fish in the basket, detailed spokes and seat spring. The UI offers free orbit, follow, side-tracking, cinematic, and pelican POV cameras, plus speed, cadence, gear, and distance readouts.

Ethan Mollick's Flooded Neo-Gothic Tower Shader

Ethan Mollick reused his Fable 5.1 prompt [5] to generate a submerged neo-Gothic tower scene with tall structures, narrow canals, and foggy vistas. Shaders compute pixel color and lighting; such demos test a model's ability to translate visual descriptions into code. Mollick praises early output but notes Claude's tendency toward dense text persists despite claimed expression improvements; real-world fit depends on individual writing and development tasks.

Summary

Both updates are substantive. GPT-6 Sol delivers stronger coding and computer-operation abilities at a lower price tier. Opus 5.5 improves long-task efficiency, cost, and output clarity. For practitioners who use AI daily, less waiting, fewer iterations, and lower costs are compelling upgrade reasons.

References

[1] OpenAI release notes: https://openai.com/index/introducing-gpt-6-sol-and-luna/

[2] Official caching dashboard: https://openai.com/index/better-prompt-caching-for-gpt-6/

[3] Anthropic Opus 5.5 positioning: https://www.anthropic.com/claude-opus-5-5

[4] Addy Osmani Three.js pelican: https://x.com/addyosmani/status/2102436416437580159

[5] Ethan Mollick tower shader: https://x.com/emollick/status/2102441628661080384