Opus 5.5 Becomes AINews' Default Model as Every Lab Cuts Prices 40-50%

Latent Space · rss · 2026-09-23

Latent Space's AINews breaks down the Claude Opus 5.5 launch and an industry-wide price war:

Launch highlights

Anthropic shipped Opus 5.5, claiming Fable 5.1-level performance at 40% lower running cost and 30% faster than Opus 5; OpenAI launched GPT-6 Sol and Luna about an hour later at 50% below GPT-5.6 pricing. Latent Space benchmarked it on its own newsletter and is migrating to Opus 5.5 immediately.
Writing improvements: puts key info up front and follows user style rules. Session limits up 20%, banked rate-limit resets for Pro/Max/Team, now default in Claude Code and the app; Sonnet 5.5 and Haiku 5.5 follow in weeks.
First Opus with Fable 5.1-class cyber/bio/frontier safeguards.

Pricing and token economics

List price cut 20% ($5/$25 → $4/$20 per 1M tokens), but higher token usage partially offsets it: at max effort, $5.98/task vs $5.86 for Opus 5 — the "40% cheaper" claim holds only at medium effort.
Cline: beats Fable 5.1 on the Intelligence Index at 2.5x lower cost.

Evals and anomalies

Third-party results: FrontierSWE 62.3% (#2 behind GPT-6 Astra's 65.5%), CursorBench 57.8% (new #1), ParseBench tables 93.9% — though charts/layout remain weak and OCR cost is deemed too high for production.
Counterintuitive: xhigh effort costs 2.8x more than medium yet scores 3.2 points lower, attributed to scope creep penalized by agentic coding benchmarks.

System card

First lab report of scaling to 100 parallel agents; ProgramBench author OfirPress flagged that its near-100% solve rate comes from a 166/200 subset excluding the hardest programs.

More from Companies & People