• Bottom line badge: Use it for cost-sensitive workloads, Wait on maximum-effort reasoning agents — the one number that decides it is 128,000, the output-token ceiling Opus 5.5 reportedly exceeded silently.
• Anthropic priced Claude Opus 5.5 at $4/$20 per million input/output tokens, a 20% list-price cut Anthropic says yields roughly 40% lower cost on typical workloads.
• OpenAI priced GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50 per million tokens, 50% and up to 58% below GPT-5.6 rates — standard pricing on OpenAI's own page, though neither company's materials state whether the cuts are permanent or introductory.
• Capability claims on both launch pages — Anthropic's 680,000-line migration and 39/40 success rate, OpenAI's factuality-eval gains — are vendor-self-reported; the OpenAI figures could not be independently reproduced by this desk.
• Independent developer Simon Willison reported Claude Opus 5.5 exceeded its 128,000-token output limit and returned no response at maximum thinking effort on a routine test — unresolved and unaddressed in Anthropic's materials as of publication.

Anthropic and OpenAI cut prices on their mid-tier and flagship-adjacent models within about an hour of each other on September 22, 2026. Claude Opus 5.5 landed at $4/$20 per million input/output tokens, a 20% list-price cut that Anthropic says works out to roughly 40% lower cost on typical workloads, while OpenAI dropped GPT-6 Sol to $2/$10 and GPT-6 Luna to $0.10/$0.50, cuts of 50% and up to 58% off GPT-5.6 pricing. Neither company's pricing page states whether the new rates are permanent or introductory. Bottom line: Use it for cost-sensitive routing, but one early tester's account of Opus 5.5 silently overflowing its output limit under heavy reasoning is a reliability question neither company's launch post addresses.

What Actually Changed

The numbers are consistent across each company's own announcement. Anthropic's Claude Opus 5.5 launch page prices the model at $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5/$25, with cached-input reads cut 60% to $0.20 per million tokens. OpenAI's GPT-6 Sol and GPT-6 Luna pages price the pair at $2/$10 and $0.10/$0.50 per million tokens respectively, down from GPT-5.6 Sol's $4/$20 and GPT-5.6 Luna's $0.20/$1.20 — listed on OpenAI's own pricing page without any promotional or expiration language attached.

The Performance Claims Are Vendor-Reported, Not Independently Verified

Every capability claim attached to this launch traces back to the vendor, not a third party. Anthropic says Opus 5.5 completed a 680,000-line code migration in under a day and succeeded 39 of 40 times on a page-load-optimization task where Opus 5 altered app behavior — figures published on Anthropic's own launch page and system card. OpenAI says GPT-6 Luna improved 5.4 percentage points on an internal factuality eval at 58% lower cost per task, and that GPT-6 Sol roughly halves GPT-5.6 Sol's error rate on the same internal benchmark; these OpenAI figures come from OpenAI's own announcement and could not be independently reproduced by this desk. None of these figures ship with a published, reproducible methodology — reported, not confirmed.

The Reliability Wrinkle Neither Launch Post Mentions

Independent developer Simon Willison, testing both releases the same day, reported that Claude Opus 5.5 at its maximum thinking/effort setting exceeded its 128,000-token output limit while reasoning through a routine SVG-generation test and failed to return any response at all. That is a single documented incident from one tester, not a benchmark — but it lands on the exact axis, maximum-effort reasoning, that agentic coding workflows lean on hardest, and it is the kind of failure mode a vendor's own launch page has no incentive to surface.

Who Should Not Switch Yet

Teams running high-volume, cost-sensitive tasks — classification, extraction, short-form generation, chat support — get an unambiguous win: GPT-6 Luna and GPT-6 Sol are priced 50-58% below their GPT-5.6 predecessors on OpenAI's own page, with vendor-reported quality gains, and Opus 5.5's $4/$20 rate undercuts Opus 5 by a fifth. Teams whose production agents run at maximum reasoning effort on long-horizon coding or research tasks should hold their current model in place until the output-limit failure is reproduced, explained, or patched — a silent non-response on a reasoning-heavy call is worse than a slow one.

The Bottom Line

Bottom line: Use it for the price cut, not yet for a blind swap of your highest-stakes reasoning workloads. The pricing is real and verifiable on both companies' own product pages, though neither page states whether the new rates are locked in for the long term or could move again, the way OpenAI's outgoing GPT-5.6 pricing reportedly was headed for a scheduled increase. The one number that decides it is 128,000 — the output-token ceiling Opus 5.5 was reported to exceed silently at maximum effort. Until that is independently reproduced or addressed, route new spend to the cost savings on already-validated task types and keep your highest-reasoning production paths on a model you've stress-tested.