OpenAI cut GPT-6 Sol and Luna prices to half of the prior generation. Anthropic, on the same day, dropped Claude Opus 5.5 task costs by 40%. Both played on the same day — but what they're really competing on isn't the models themselves; it's per-task unit cost.

What this is

Last week, OpenAI released two "lite" versions — GPT-6 Sol and Luna — with prices halved directly against prior-generation GPT-5.6: input dropped from $4/million tokens to $2, output from $20 to $10. The cheaper Luna runs at just $0.10 input / $0.50 output. Anthropic fielded Claude Opus 5.5, with typical task costs down roughly 40% versus Opus 5, and output speeds up over 30%.

Both companies are loudly promoting the same metric this round — per-task pricing (cost per completed task, not per token). OpenAI's published comparisons: GPT-6 Sol beats Claude Opus 5 on AutomationBench (a cross-application enterprise workflow test) at just 9% of Opus 5's per-task cost; Luna hits equivalent performance at roughly 1% of the prior-generation Sol's cost. In plain terms: applications that were previously affordable can now confidently scale.

Sam Altman framed it on social media as "democratization of intelligence." Translated into industry terms: frontier models are no longer priced as luxury goods — they're being sold as fast-moving consumer goods.

Industry view

The developer community broadly welcomes this — especially for Coding Agents (automated assistants that write code for you) and long-task scenarios. Experiments previously shelved due to compute bills now have room to be re-run. OpenAI also raised Prompt Caching's default hit rate, giving cached input tokens a 90% discount — effectively another red packet for high-frequency callers. With compute getting cheaper, Agent-class business models will be rebuilt from the ground up.

But the objections deserve to be put on the table:

  • Benchmark scores ≠ real-world workloads. AutomationBench and OSWorld — both published by the vendors — are test sets they selected themselves. Multiple enterprises have reported that failure rates are noticeably higher than benchmark scores suggest, when models are dropped into real long-chain, cross-system business workflows.
  • "Halved" is marketing language — halved from what, exactly? OpenAI is benchmarking against its own GPT-5.6 promotional price, not against Anthropic's new model. Anthropic hasn't published a granular cost comparison against Sol either. The real magnitude of the cuts needs independent measurement.
  • Cheaper means deeper lock-in. Both are cutting prices, but the models remain closed APIs (callable only through the vendor's interface, with no open underlying weights). Once business is deeply embedded into one vendor's pipeline, the future cost of price hikes, rate limits, or service terminations will be higher than ever. "Model substitutability" is about to become the next agenda item for enterprise IT.

Impact on regular people

For enterprise IT: The framing of this year's AI budget needs to change. Stop asking "which model is strongest" — start asking "how much per task, and can we push it below a target threshold." A new column will be added to procurement evaluation tables.

For individual professionals: You can safely increase how often you use AI tools. Features that previously got rationed due to token cost anxiety (letting AI read an entire PDF, multi-round comparisons, batch rewriting) are no longer cost-blocked. The gap between "knows how to use it" and "doesn't" will be eclipsed by the gap between "uses a lot" and "uses a little".

For consumer markets: AI-embedded products will get a stealth upgrade. SaaS products (subscription software) that previously bragged "powered by GPT-4" will quietly swap Sol in underneath without telling you — but the capability ceiling will be quietly raised. You'll notice "this thing feels noticeably better than it did six months ago."