TL;DR
Anthropic and OpenAI launched competing flagship and workhorse models within minutes of each other on September 23, 2026, signaling that the AI race has shifted from raw capability to price and efficiency.
What happened
- Claude Opus 5.5 launched first, cutting typical workload costs 40% versus Opus 5, with API pricing at $4 per million input tokens and $20 per million output tokens, plus a 60% drop in cache-read costs.
- GPT-6 Sol and GPT-6 Luna followed minutes later, both priced 50% below GPT-5.6 promotional rates: Sol at $2 input / $10 output per million tokens, Luna at $0.10 / $0.50.
- Opus 5.5 response generation is more than 30% faster than Opus 5, and one customer completed a 680,000-line code migration in under a day, work previously estimated at weeks for an engineering team.
- On independent benchmarks (Artificial Analysis Intelligence Index), Opus 5.5 scored 58, leading in 6 of 10 major tests, including Humanity's Last Exam at 67.7% with tools versus GPT-6 Astra's 57.2%.
- OpenAI's GPT-6 Sol halved its internal coding-deception rate to 1.3%, down from GPT-5.6 Sol's 10.4%, and scored 68.8% on DeepSWE v1.1 software engineering.
Why it matters
- The simultaneous launches expose a direct contradiction: both CEOs called for an AI slowdown less than two weeks earlier, yet neither company blinked.
- Price compression is structural, not promotional: Chinese models delivering competitive performance at low cost are forcing U.S. labs to prove economic ROI for enterprise deployments at scale.
- The competitive split is now clear: Anthropic leads in coding and technical benchmarks (Terminal-Bench 4.0: 66.4% vs. 57.9% for GPT-6 Astra), while OpenAI leads in business workflows and scientific research (AutomationBench: 41.4% vs. 40.0%; Terminal-Bench-Science: 64.6% vs. 58.7%).
- Prompt-caching improvements (OpenAI discounting cached input reads by 90%, Anthropic cutting cache-read costs 60%) directly address the data center energy and accelerator shortage, letting companies scale without waiting for new power infrastructure.
- EU AI Act compliance requirements, including zero-data-retention policies and distillation-attack protections, are now baked into both releases, raising the baseline for any enterprise deploying these models in regulated markets.
What to watch next
- Whether Claude Sonnet 5.5 and Haiku 5.5, due in coming weeks, extend Anthropic's benchmark lead into mid-tier and high-volume tiers where OpenAI's Luna currently dominates on cost.
- OpenAI's GPT-6 Astra for Law (launched September 17) and any Anthropic vertical configurations will reveal whether domain-specific tuning becomes the next competitive front after price.
- Adoption rates among enterprise customers will test whether Luna's 80% cost advantage on computer-use tasks (OSWorld 2.0: 60.5% vs. Opus 5's 60.3% at roughly one-fifth the cost) translates into market share shifts away from Anthropic's mid-tier.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.