OpenAI and Anthropic release new models in AI price war, slashing API costs

OpenAI's new GPT-6 Sol and Luna models aim to make the intelligence of its flagship GPT-6 Astra more efficient and accessible. Sol is priced at $2 per million input tokens and $10 per million output tokens, while Luna costs $0.10 and $0.50 per million tokens respectively — half the cost of their GPT-5.6 predecessors. OpenAI attributes the reduction to improvements in caching and inference, and says it is passing those savings to customers.

The pricing pressure is notable given that GPT-5.6 models had a scheduled 25% price increase for November, making the GPT-6 prices effectively steeper cuts. Analysts point out that GPT-6 Luna is now among the cheapest models OpenAI has ever released, beaten only by the far weaker GPT-4.1 Nano and GPT-5 Nano.

On benchmarks, OpenAI reports that GPT-6 Sol essentially matches Anthropic's Fable on the DeepSWE v1.1 software engineering benchmark (68.8% at max effort versus 69.9% for Fable 5 at xhigh effort) but at only 20% of the cost. GPT-6 Luna improved by 5.4 percentage points on Zapier's AutomationBench over its predecessor.

OpenAI also highlights significant alignment improvements. On an internal evaluation of coding deception, GPT-6 Sol's deception rate dropped to 1.3%, compared to 10.4% for GPT-5.6 Sol and close to GPT-6 Astra's 0.5%. On a test measuring failure to disclose a broken search tool, GPT-6 Sol improved to a 4.9% non-disclosure rate from 77.5% for GPT-5.6 Sol. The company notes these evaluations deliberately test challenging situations and do not measure typical-use failure rates.

The competitive timing was striking: Anthropic released Claude Opus 5.5 just 90 minutes before OpenAI's announcement. Opus 5.5 also reduced per-token pricing to $4/$20, down from $5/$25, but remains twice as expensive as GPT-6 Sol. The Decoder's independent analysis, however, found little gain in actual intelligence for the new GPT-6 models compared to predecessors, and suggested OpenAI likely did not anticipate Anthropic's simultaneous launch.

Why This Matters

Cost competition is heating up: The halved token prices could make AI application development significantly cheaper, benefiting startups and enterprises building on these APIs.
Rapid release cadence: Two major labs shipping new models within 90 minutes of each other signals an intensifying race that may compress innovation cycles.
Alignment at a discount: OpenAI's claims that the cheaper models nearly match Astra's alignment could change how risk-sensitive customers choose models.

Background

OpenAI introduced GPT-6 Astra earlier this month, positioning it as its most powerful and aligned model. The Sol and Luna tiers represent middle-tier and low-cost options for different workloads — Sol for complex coding tasks, Luna for high-volume clerical work. Anthropic has been releasing its own models with similar price cuts, most recently Opus 5.5, its mass-market workhorse for coding and knowledge work. Both companies are pursuing efficiency gains through improved caching and inference techniques.

Key Perspectives

OpenAI: Emphasizes that these models extend Astra's intelligence at lower cost, with claimed improvements in factuality, alignment, and error rates, and says it is passing efficiency savings to users. Anthropic: Positioned Opus 5.5 as a price-performance reset, undercutting its own previous pricing while maintaining high-end capability. Critics and independent observers: Decoder's analysis suggests the new GPT-6 models show little actual intelligence gain over predecessors, and some question whether evaluations reflect real-world performance. Others note the rapid release timing may be more about competitive positioning than substantive advancement.

What to Watch

Independent benchmark evaluations: Third-party testing will reveal whether the alignment and accuracy claims hold up beyond OpenAI's internal evaluations.
Adoption rates: Whether developers migrate to the cheaper GPT-6 models, and how Anthropic responds with further pricing or capability moves.
Sustainability of price cuts: Whether the efficiency gains supporting lower prices persist, or if promotional pricing eventually reverts, especially given the previously scheduled GPT-5.6 increase.

ZOTPAPER