Reviewed and written up.
Added Sep 22, 2026
Claude Opus 5.5 is Anthropic's September 22, 2026 update to Opus 5, priced at $4 per million input tokens and $20 per million output (down from $5/$25), with cache reads cut from $0.50 to $0.20 per million and cache writes from $6.25 to $5. A faster mode in Claude Code and the Claude Platform runs up to 2.5x quicker at $8/$40 per million tokens. Anthropic's own benchmarks put it ahead of the larger Claude Fable 5.1 on agentic coding (66.4% vs 55.8% on Terminal-Bench 4.0), computer use (81.8% vs 80.7% on OSWorld 2.0), and knowledge work (1846 vs 1735 Elo on GDPval-AA v2.1), while generating output more than 30% faster than Opus 5. Five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans were also increased alongside the release.
Why: It is priced 20% below Opus 5 on the headline rate and roughly 40% cheaper on typical agentic workloads once cache-read savings are counted, while benchmarking ahead of the more expensive Fable 5.1 on the coding and computer-use tasks this directory's readers actually run. That combination, cheaper than its own predecessor and ahead of the tier above it, is unusual enough to be the pick over Opus 5 for anyone already on Claude.
Added Sep 22, 2026
GPT-6 Sol is OpenAI's September 22, 2026 update to GPT-5.6 Sol, priced at $2 per million input tokens and $10 per million output, down from $4/$20, a permanent price cut OpenAI has confirmed is not promotional. It is positioned for complex, repeated developer work: building features, reviewing code, debugging, and analyzing data. On the DeepSWE v1.1 software-engineering benchmark, independent testing found Sol reaches 68.8% at maximum effort for $2.14 per task, against Claude Fable 5's 69.9% at its highest effort setting for $21.63 per task, near-identical accuracy at roughly a tenth of the cost. OpenAI says it makes about half as many mistakes as GPT-5.6 Sol.
Why: The near-parity DeepSWE result against Claude Fable 5 at a tenth of the cost is the reason to look at this model specifically for coding workloads: it is a real price war entry, not a token repricing with nothing behind it, though independent reviewers note overall intelligence scores are flat versus GPT-5.6 rather than improved.
Added Sep 21, 2026
Grok 4.7 is SpaceXAI's flagship model, released September 21, 2026, for coding, agents and knowledge work. It takes text and images and returns text, with a 500,000-token context and a knowledge cutoff of May 2026. Reasoning effort is low, medium, high or xhigh, and high is the default. The API name is grok-4.7. xAI prices prompts under 200,000 tokens at $2 per million input tokens, $0.50 for cached input, and $6 per million output tokens. A request that reaches 200,000 tokens is billed at double those rates for every token. A faster serving variant costs twice the standard rates and is limited to Cursor and Grok Build.
Why: Grok 4.5 is not the model xAI tells developers to call anymore. 4.7 is the current flagship, with image input and a published price that matches what xAI charged for Grok 4.6.
Added Sep 22, 2026
GPT-6 Luna is OpenAI's September 22, 2026 update to GPT-5.6 Luna, priced at $0.10 per million input tokens and $0.50 per million output, down from $0.20/$1.20. It is positioned for high-volume, tightly-defined tasks with a clear goal: summarizing documents, extracting information, and answering straightforward questions, rather than open-ended reasoning or coding. Independent testing on the DeepSWE v1.1 benchmark found Luna matches GPT-6 Sol's maximum-effort score for just $0.22 per task.
Why: At $0.10/$0.50 per million tokens it is priced for workloads where cost per call is the entire decision, and matching Sol's benchmark score at a fraction of Sol's price on at least one measured task makes it worth trying before defaulting to a more expensive tier for simple, high-volume work.