- OpenAI GPT-6 Sol (Sep 22, 2026): $5 input, $30 output, 1.05M context. Same $5/$30 tier as Opus 5.5 but with cheaper cached input at $0.20 per million. Best for new agentic-coding workflows.
- Anthropic Claude Fable 5.1 (Sep 1, 2026): $10 input, $50 output, 1M context. Anthropic's top tier above Opus 5.5; for the hardest research questions where cost is secondary.
Models at the other end - GPT-6 Luna, Claude Haiku 4.5, and Grok 4.7 - are cheaper and faster but lower-capability substitutes for everyday high-volume work.
What to do today
Three actions for the September 23 decision window:
- If you were using Opus 5 (the prior tier), migrate to Opus 5.5 by changing the model id to `claude-opus-5-5`. Pricing drops 40%; behaviour is largely the same with adaptive thinking now on by default.
- If you were using GPT-5.5 Pro for everyday coding, route the same traffic to Opus 5.5 at roughly 7.5x cost savings and accept Opus 5.5's adaptive thinking as the new default reasoning strategy.
- If you are choosing between the two models for a new deployment, start with Opus 5.5 fast mode for interactive work and add GPT-5.5 Pro for explicit max-effort research where the cost is justified.
The Anthropic migration guide for Opus 5.4 → Opus 5.5 is published on docs.anthropic.com; the change is non-breaking for most workloads.
Frequently Asked Questions
Is Opus 5.5 better than GPT-5.5 Pro for coding?
On price-performance, Opus 5.5 wins for everyday agentic coding - you get within a few points of GPT-5.5 Pro on Terminal-Bench 2.0 at less than 15% of the per-token cost.
On raw accuracy at the very hardest coding benchmarks, GPT-5.5 Pro at xhigh still leads (Anthropic, September 2026; OpenAI, April 2026).
Does Opus 5.5 have cached input pricing?
Yes - cached input reads on Opus 5.5 cost $0.20 per million tokens, 5% of the standard $4 input rate.
GPT-5.5 Pro does not have a cached-input discount tier. For workflows that benefit from prompt caching, the Opus 5.5 effective price drops further than the headline 7.5x gap (Anthropic pricing, September 2026).
Can I use the OpenAI Chat Completions endpoint with GPT-5.5 Pro?
No - GPT-5.5 Pro only ships on the OpenAI Responses API and Batch API.
The Chat Completions endpoint serves GPT-5.5 (non-Pro) and the GPT-6 series. Customers needing the Chat Completions interface with max-effort reasoning should use GPT-6 Sol at the same $5/$30 tier as a substitute (OpenAI API docs, April 2026).
Which model has the better knowledge cutoff?
Opus 5.5 inherits Claude's 2026 training scope. Anthropic publishes model cards with the cutoff at training time.
GPT-5.5 Pro was trained with a December 1, 2025 cutoff per the OpenAI API docs - five months older than the typical Anthropic cutoff for Opus-class models. For time-sensitive factual tasks, Opus 5.5 has the newer knowledge base (OpenAI API docs, April 2026; Anthropic, September 2026).
Does Opus 5.5 support voice input?
No - Opus 5.5 is text-in / text-out like most frontier LLMs. For voice workflows, both Anthropic and OpenAI ship voice-specific models. Anthropic routes through Claude's SDK; OpenAI offers the Realtime API for low-latency voice. The frontier text models are unrelated (Anthropic, September 2026).
Is GPT-5.5 Pro being deprecated?
Not as of September 2026 - GPT-5.5 Pro remains OpenAI's established max-effort tier.
The GPT-6 series (Astra, Sol, Luna) sits one tier lower in pricing and one tier newer in training. For workloads where GPT-5.5 Pro's reasoning profile is the established fixture, the model continues to ship (OpenAI API docs, April 2026; aireleasetracker.com, September 2026).