The AI Price War Just Got Brutal — And It's Not About Raw Power Anymore
Anthropic and OpenAI dropped heavily discounted models this weekend, each claiming efficiency wins over the other's flagship. The race has shifted from benchmark chasing to cost-per-task domination.
The weekend that redefined the mid-tier AI market
Two companies released new model tiers on the same day. Anthropic led with Claude Opus 5.5, OpenAI countered with GPT-6 Sol and GPT-6 Luna. Both pitched the same story: you can now get near-flagship performance at roughly half the cost. What happened next reveals how far the AI arms race has drifted from pure capability toward economic warfare.
Anthropic’s efficiency play
Claude Opus 5.5 arrived September 22 across the Claude Platform, AWS, Google Cloud, and Microsoft Azure. Anthropic’s claim is blunt. In typical workloads, it performs comparably to Claude Fable 5.1, the company’s top-tier model, while costing 40 percent less to run. The API pricing backs that up: input at $4 per million tokens, output at $20 — a 20 percent drop from Opus 5. Cached reads fell to $0.20 per million, down 60 percent.
The benchmark numbers tell a mixed story. On Terminal-Bench 4.0, which measures agent-style coding, Opus 5.5 scored 66.4 percent, ahead of Fable 5.1’s 55.8 percent and OpenAI’s GPT-6 Astra at 57.9 percent. But on AutomationBench, which tracks business workflow tasks, and Terminal-Bench-Science 0.1, GPT-6 Astra pulled ahead. The model is sharper at coding than reasoning through complex operational sequences.
Anthropic also highlighted safety improvements, claiming an internal audit record best ever and an 85 percent reduction in attempts to bypass constraints compared with Opus 5. The thinking mode — the slow, deliberate reasoning step some users rely on — is no longer available for Opus 5.5, a move that underscores the trade-off Anthropic is making: speed and cost over的深度 reasoning for this tier.
CEO Dario Amodei has been publicly urging the industry to slow the pace of AI development. Opus 5.5 is his first release since that stance, and the pricing cut suggests he’s betting that restraint and efficiency will win customers, not just raw speed.
OpenAI’s counter-strike
OpenAI moved on the same day, launching GPT-6 Sol and GPT-6 Luna. Both were trained using the same approach as GPT-6 Astra, the company’s premium model announced earlier in September, but optimized for lower cost and faster inference.
The discount is aggressive. GPT-6 Sol runs at $2 in and $10 out per million tokens — a 50 percent cut from GPT-5.6 Sol. GPT-6 Luna drops to $0.10 in and $0.50 out, down from $0.20 and $1.20. These are not marginal adjustments. They restructure the unit economics of running AI at scale.
OpenAI claims GPT-6 Sol makes roughly half the mistakes of GPT-5.6 Sol and closes the alignment gap with Astra. On DeepSWE v1.1, a benchmark measuring practical software development, Sol scored 68.8 percent compared with Claude Fable 5’s 69.9 percent — a gap of 1.1 points at about 80 percent lower cost per task. On AutomationBench, Sol posted 33.2 percent, beating Claude Opus 5’s 26.9 percent while costing only 9 percent as much.
Availability mirrors Anthropic’s strategy of broad deployment. ChatGPT Work and Codex gained access immediately for Plus, Pro, Business, Enterprise, and Edu subscribers. Free and Go users can access GPT-6 Luna on desktop. The API identifiers are gpt-6-sol and gpt-6-luna. Notably, the standard ChatGPT interface does not yet include these models.
What shifts when price becomes the weapon
The immediate effect is a compression of margins for everyone except the two largest players. Anthropic and OpenAI can absorb these cuts because their infrastructure scale lets them push cost per token down across every tier. Smaller labs cannot match this without sacrificing margins they do not have.
Enterprise buyers benefit most in the short term. The cost per million tokens on agent coding and workflow automation has dropped sharply. Teams running large volumes of API calls will see their bills shrink, and that changes deployment decisions — more tasks move from batch processing to real-time pipelines because the economics now justify it.
Developers should expect a brief period of instability as companies update their model routing logic. The benchmark rankings are close enough that small shifts in prompt design or caching strategy may matter more than the model choice itself. Anthropic leads on Terminal-Bench; OpenAI leads on AutomationBench. The right model depends on what you are actually building.
The deeper signal
Both companies are positioning the 5.5 and 6 Sol tiers as the new default. The flagships — Fable 5.1 and GPT-6 Astra — remain important for research and frontier work, but the commercial battlefield has moved downward. This is what maturity looks like in a product cycle that barely existed two years ago.
Anthropic’s safety metrics and Amodei’s public call for caution give the company a differentiation angle. OpenAI’s price cuts and benchmark lead on workflow tasks give it a commercial one. The next few weeks will show whether customers reward the caution narrative or the cost narrative. Claude Sonnet 5.5 and Haiku 5.5, arriving in coming weeks, and OpenAI’s follow-on releases will clarify which direction the market is taking.
The models are fast, the prices are low, and the competition is real. The only question left is who pays when the race stops slowing down.