OpenAI fires back: GPT-6 Sol at half price

Hours after Anthropic dropped Claude Opus 5.5 on September 22, 2026 (UTC-4), OpenAI fired back with GPT-6 Sol and GPT-6 Luna — two new models in the GPT-6 family that cut API prices by 50% compared to their GPT-5.6 predecessors. The same-day release wasn't subtle: the two biggest frontier labs are now in a full-on price war, and developers are the winners.

What shipped

GPT-6 Sol is the mid-tier workhorse, priced at $2/$10 per million input/output tokens — half the $4/$20 of GPT-5.6 Sol. GPT-6 Luna is the budget option at $0.10/$0.50, down from $0.20/$1.20. GPT-6 Astra, released September 3, remains the flagship at $10/$50.

Both models use the same training methods as Astra, bringing its professional work, factuality, coding, computer use, and alignment advances to cheaper, faster tiers. OpenAI also improved prompt caching across GPT-6, delivering 90% discounts on cached input reads and higher default cache hit rates. GitHub reports these caching changes reduced fresh-processing tokens by over 50% across billions of requests.

Availability is rolling out now: ChatGPT Work and Codex get Sol and Luna for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access Luna in the desktop app. Both models are live in the API as gpt-6-sol and gpt-6-luna. They're not yet available in the standard Chat interface.

The benchmark picture

OpenAI's comparison frames Sol as punching above its price class against Claude's lineup.

Source: OpenAI official announcement, September 22, 2026. Competitor scores from publicly available reports; Fable 5.1 AutomationBench cost omits Opus 5 fallback charges (~40% of tasks). "—" indicates the model was not included in that benchmark by OpenAI.

The headline number: on AutomationBench, Sol at xhigh beats Opus 5 at max effort while costing 91% less per task. On DeepSWE, Sol comes within 1.1 points of Fable 5's top score at roughly 80% lower cost. On OSWorld, Sol matches Opus 5 at 80% lower cost. These are not marginal improvements — they're order-of-magnitude shifts in the cost-performance curve.

Factuality also improved: Sol makes about half as many factual mistakes as GPT-5.6 Sol, approaching Astra-level reliability. Luna at higher effort matches GPT-5.6 Sol at roughly one-hundredth the cost.

Why it matters

The timing is the story. Anthropic released Opus 5.5 at $4/$20 — already 20% cheaper than Opus 5 — and OpenAI responded the same day with Sol at $2/$10. That's half of Anthropic's new price for a model that, on OpenAI's benchmarks, matches or beats Opus 5 on most tasks. The frontier model pricing assumption of $5+/M input and $25+/M output is dead. The new ceiling for "good enough for most work" is $2/$10, and the floor is $0.10/$0.50.

For developers, this changes unit economics. A coding agent that cost $3 per task on Opus 5 now costs $0.27 on Sol. OpenAI's own researchers burn $600/day in API tokens at the median and $7,000/day at the 90th percentile — at Sol prices, that drops to $300 and $3,500. Startups building agentic products can run 10x more iterations for the same budget, which accelerates the entire application layer.

The three-tier structure — Astra ($10/$50), Sol ($2/$10), Luna ($0.10/$0.50) — is a deliberate segmentation. Astra handles the hardest, highest-stakes work. Sol covers everyday professional and coding tasks. Luna serves high-volume, low-complexity use cases. This mirrors Anthropic's Fable/Opus/Sonnet/Haiku lineup but with sharper price points. The question is whether Sol is "good enough" to cannibalize Astra usage — if it is, OpenAI's revenue per token drops even as volume explodes.

The critical lens

Let's read the benchmarks carefully. OpenAI's comparison uses Sol at xhigh or max effort against competitors at their own max — but effort level affects both score and cost. On AutomationBench, Sol at xhigh scores 33.2% at $0.27/task. The same benchmark from Anthropic's Opus 5.5 announcement shows Opus 5.5 at 40.0% and GPT-6 Astra at 41.4%. OpenAI doesn't include Opus 5.5 in its table — because Opus 5.5 didn't exist when the benchmarks were run, but also because including it would complicate the "Sol beats Claude" narrative. The two labs are now publishing competing benchmark tables with different model sets and effort levels, making direct comparison a minefield.

The 50% price cut is real but needs context. GPT-5.6 Sol launched at $5/$30 in June, then dropped to $4/$20 as a promotional price. The new $2/$10 is 50% off the promotional price, but 60% off the original launch price. OpenAI is framing it as a 50% cut from the most recent price, which is technically accurate but undersells how far prices have fallen in three months.

And the same-day release pattern deserves scrutiny. Anthropic announced Opus 5.5 in the morning (US time); OpenAI announced Sol and Luna in the evening. Whether this was coordinated scheduling or competitive timing, the effect is the same: the market now expects simultaneous releases and price matching. The "pacing the frontier" narrative from Amodei's September 12 essay looks even more hollow when both labs are dropping models within hours of each other.

What to watch

Watch Anthropic's response. Opus 5.5 at $4/$20 is now twice as expensive as GPT-6 Sol at $2/$10. If Anthropic's benchmarks hold up — Opus 5.5 scores 66.4% on Terminal-Bench 4.0 vs. Sol's likely lower score — the price premium may be justified for hard coding. But for general professional work, Sol's cost advantage will pull developers toward OpenAI. Expect Anthropic to either cut Opus 5.5 pricing or emphasize the benchmark gap in the coming weeks.

Watch ChatGPT Work adoption. Sol and Luna launch in ChatGPT Work and Codex but not in the standard Chat interface. This is a deliberate push toward the work-focused product, positioning Work as the professional alternative to ChatGPT's consumer chat. If Work usage spikes, it validates OpenAI's bet that agentic work — not casual chat — is where monetization grows.

Watch the Luna tier. At $0.10/$0.50, Luna is cheaper than most open-source models when you factor in hosting costs. If Luna's quality is solid, it could kill the economic case for self-hosting 7B-13B parameter models for production workloads. That's the real disruption: not beating the frontier, but making the cheap tier good enough to commoditize the middle.