OpenAI just halved GPT-6 prices. Read the fine print before you celebrate.

The cut is real. The baseline is doing some work.

On 22 September OpenAI released GPT-6 Sol and GPT-6 Luna, the workhorse and budget tiers sitting under its GPT-6 Astra flagship, and halved their API prices. Sol drops from $4 to $2 per million input tokens and from $20 to $10 for output. Luna falls from $0.20 to $0.10 on input and from $1.20 to $0.50 on output. The models are live in the API as gpt-6-sol and gpt-6-luna, and in ChatGPT Work and Codex for paid tiers, with Luna reaching free users through the desktop app.

Look closely at how OpenAI states the discount, though. The launch post says prices are down 50% "compared with their GPT-5.6 promotional pricing". The baseline for the halving was itself a promotion. That does not make the new numbers less real: $2 and $10 is what your invoice will say, and OpenAI presents the rates as standing prices rather than a launch special. But "we halved prices" is marketing work layered on a long industry trend, because headline per-token prices have been falling almost continuously since the first GPT-4 rate cards. The news is that the fall got steeper, not that generosity broke out.

Every chart in the launch post grades its own homework

OpenAI leans on four evaluations. On AutomationBench, a business-workflow test, Sol at its highest effort setting scores 33.2% at $0.27 per task, which OpenAI says beats Claude Opus 5 at max effort at roughly a ninth of the cost. On Agents' Last Exam, Sol posts 56.4%. On DeepSWE, a software-engineering eval, Sol reaches 68.8% and Luna 66.6%. On OSWorld's offline computer-use suite, Sol edges Claude Opus 5 by a fifth of a percentage point at what OpenAI says is about 80% lower cost per task.

All of these numbers are vendor-run, and all of the comparisons target Anthropic models. OpenAI's own footnote quietly concedes that the sharpest one is not clean: the cost figure shown for Claude Fable 5.1 "understates its actual cost" because it leaves out fallback calls to Opus 5 that happened on roughly 40% of tasks. Credit for the disclosure. A footnote still does not fix a chart.

The factuality claim deserves the same squint. "About half as many mistakes" as GPT-5.6 Sol comes from an internal evaluation built on conversations where users had already flagged an error, a sample OpenAI itself says is not representative of everyday use. Fewer errors on a deliberately hostile set is a genuine signal. It is not a warranty.

Anthropic answered the same day

The Claude models OpenAI chose as punching bags stopped being Anthropic's best before the day ended. On 22 September Anthropic released Claude Opus 5.5 at $4 input and $20 output per million tokens, output down from $25, and says it "performs at the level of Claude Fable 5.1 on most work" while costing about 40% less than Opus 5 on typical workloads. The launch page carries its own sheaf of self-run benchmark wins and customer testimonials, plus a claim of output generation more than 30% faster than Opus 5. Anthropic says Sonnet 5.5 and Haiku 5.5 follow within weeks, so the repricing wave has at least two more crests coming.

The upshot: in one afternoon, the price of serious machine intelligence fell at both major API shops at once. That is not a coincidence. That is a price war, and a price war is the most reliably good news a buyer ever gets.

The quiet story is caching

The rate card is only half the bill for agent workloads. OpenAI paired the cuts with improved default prompt caching and a 90% discount on cached input-token reads, plus a dashboard, diagnostics and explicit cache breakpoints for tuning. It cites GitHub reporting that caching improvements cut the share of prompt tokens needing fresh processing by more than half across billions of requests. Anthropic pulls the same lever from the other side: Opus 5.5 cache reads cost $0.20 per million tokens.

For long-running agents that reread the same context hundreds of times, effective cost falls faster than the headline prices suggest. It also makes vendor cost-per-task charts even harder to audit from the outside, because cache behaviour depends on the workload.

OpenAI's post offers a rare glimpse of where all this is heading: valued at API prices, the median OpenAI researcher now burns more than $600 of tokens a day, and the 90th percentile passes $7,000. Sustained agent use at that intensity is what both labs are pricing for.

What a buyer should actually do

Reprice first, argue later. If you run production workloads, moving off GPT-5.6-era pricing is immediate money, and Sol's $2 and $10 list now undercuts Opus 5.5's $4 and $20. Then benchmark on your own tasks at your own effort settings, because the effort dial moves your cost more than the rate card does. Treat every cross-vendor chart from either lab as advertising until independent evaluators run Sol, Luna and Opus 5.5 side by side, which the same-day releases have made briefly impossible. And remember that published prices change without notice, which is why we log each one with a source link and a verified-on date in our pricing tracker.

Our take

The 50% figure is marketing arithmetic on a promotional baseline, and the benchmark charts grade the vendor's own homework against a rival's superseded models. Both of those things are true alongside a third: real list prices for capable models halved in an afternoon, twice, and that is the most buyer-friendly thing the frontier labs have done all quarter. Take the money. Park the capability claims until someone without a booth measures them. September also rewrote the rulebook while it rewrote the rate card: see the EU AI Act's new deadlines.