Anthropic, OpenAI, and xAI each released new API-priced models within a 48-hour window in late September 2026 — Grok 4.7 on September 21, Claude Opus 5.5 and GPT-6 Sol/Luna both on September 22, the latter two reportedly just 90 minutes apart. None of the three is purely an apples-to-apples flagship comparison: Opus 5.5 and Grok 4.7 are each company's current top general-access model, but GPT-6 Sol and Luna actually sit below OpenAI's real flagship, GPT-6 Astra, as cheaper mid- and low-tier options. This comparison sticks to each company's own published pricing and benchmark claims.
This comparison is based on official pricing pages and launch announcements from Anthropic, OpenAI, and xAI, verified as of 14 September 2026 — the same official-pricing-page standard we used comparing Claude and ChatGPT's consumer subscription plans.
What each company actually shipped
| Claude Opus 5.5 | GPT-6 Sol | GPT-6 Luna | Grok 4.7 | |
|---|---|---|---|---|
| Company | Anthropic | OpenAI | OpenAI | xAI |
| Release date | Sept 22, 2026 | Sept 22, 2026 | Sept 22, 2026 | Sept 21, 2026 |
| Input price / 1M tokens | $4 | $2 | $0.10 | $2 |
| Output price / 1M tokens | $20 | $10 | $0.50 | $6 |
| Cache read price / 1M | $0.20 | Not separately listed | Not separately listed | $0.50 |
| Context window | 1,000,000 tokens | Not specified at launch | Not specified at launch | 500,000 tokens |
| Position in company's lineup | Top general-access model | Mid-tier, below GPT-6 Astra | Low-tier, below GPT-6 Astra | Top general-access model |
Anthropic's own pricing page lists Opus 5.5 at $4 input / $20 output per million tokens — a 20% cut from Claude Opus 5's $5/$25 — and frames cache reads, which it says make up the majority of agentic and coding work costs, as the bigger change: $0.20 per million versus Opus 5's $0.50, a 60% reduction. OpenAI's own announcement prices GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50, both described as permanent prices (not promotional) and both 50% below their GPT-5.6 predecessors' promotional rates. xAI's own developer pricing page lists Grok 4.7 at an unchanged $2/$6 from Grok 4.6, despite xAI stating the new model runs on a larger, 2.1-trillion-parameter base.
Why "cheapest" depends entirely on which number you read
Read input price alone and GPT-6 Luna wins by a wide margin at $0.10. Read output price, which dominates cost on longer responses, and Luna still leads at $0.50, with Grok 4.7's $6 ahead of GPT-6 Sol's $10 and Opus 5.5's $20. But Luna and Sol are not positioned by OpenAI as competitors to Opus 5.5 or Grok 4.7 — OpenAI's own materials describe GPT-6 Astra, not Sol or Luna, as its flagship for the most demanding work. Comparing Opus 5.5 or Grok 4.7 directly against Sol or Luna on price alone compares a top-tier model against a mid- or low-tier one from a different company.
What the cross-company benchmark claims actually show
All benchmark figures below come from each company's own self-reported testing, not independent third-party evaluation, and should be read with that caveat. OpenAI's own GPT-6 Sol/Luna announcement reports GPT-6 Sol at its highest reasoning effort scoring 33.2% on AutomationBench at $0.27 per task, against a reported 26.9% for Claude Opus 5 (not Opus 5.5) at $0.27 × 11.1 per task — OpenAI's own framing, not an Anthropic figure. The same OpenAI announcement reports GPT-6 Sol scoring 60.5% on OSWorld 2.0 offline versus 60.3% for Claude Opus 5, at roughly 80% lower cost per task by OpenAI's accounting. Separately, xAI's own Grok 4.7 announcement states the model trails Claude Fable 5.1 Max — Anthropic's separate, more expensive reasoning-focused model — on several coding and long-task benchmarks xAI itself listed.
None of these self-reported comparisons tests Opus 5.5 directly against GPT-6 Sol, Luna, or Grok 4.7 using the same methodology at the same time — each company benchmarked its new model against whichever competitor model was current when its own announcement was written. The same self-reported-benchmark caveat applied when we compared Comet, Claude in Chrome, and Gemini in Chrome's respective safety claims.
Batch pricing tells a different story than live API pricing
Each company also publishes a cheaper, asynchronous batch-processing rate for non-time-sensitive work, and the discount isn't uniform. Anthropic's own pricing documentation lists Opus 5.5's batch rate at $2 input / $10 output per million tokens — exactly half the live rate — against Opus 5's batch rate of $2.50/$12.50, which is also a 50% cut from its own live price. Per OpenAI's own GPT-6 Sol and Luna documentation, both models' knowledge cutoffs differ by over a month — April 20, 2026 for Sol versus May 18, 2026 for Luna — a detail easy to miss when comparing only price and benchmark scores, but relevant for anyone relying on either model's awareness of recent events.
What each one still doesn't tell you
- Claude Opus 5.5: Anthropic's own fast-mode pricing (2.5x speed) runs $8/$40 per million tokens, double the standard rate — the headline $4/$20 price doesn't apply to every usage mode.
- GPT-6 Sol and Luna: per OpenAI's own announcement, neither model is available in the standard ChatGPT "Chat" surface at launch — both ship first in ChatGPT Work, Codex, and the API, with Luna reaching Free and Go users only through the desktop app.
- Grok 4.7: the faster variant, available only inside Cursor and Grok Build (not the public xAI API), costs double the standard token rate and is excluded from Grok Build's free tier.
Which one fits which workload
- You need the largest context window for long documents or codebases: Claude Opus 5.5 at 1,000,000 tokens, double Grok 4.7's 500,000-token window.
- You want the lowest possible per-token cost for high-volume, lower-complexity tasks: GPT-6 Luna at $0.10/$0.50, though it's explicitly a lower-tier model, not a flagship-class competitor.
- You want flagship-class output at the lowest output price among the three top-tier models: Grok 4.7 at $6 per million output tokens, versus GPT-6 Sol's $10 (mid-tier) and Opus 5.5's $20 (flagship-tier).
- Agentic or coding work dominated by cache reads: Anthropic's own materials position Opus 5.5's 60% cache-read cut as specifically targeting this cost driver, which the headline input/output prices alone don't reflect.
Frequently asked questions
Is GPT-6 Sol the same tier as Claude Opus 5.5?
No. OpenAI's own materials position GPT-6 Astra, not Sol, as its flagship model for the most demanding work; Sol and Luna are mid- and low-tier models built for cheaper, high-volume use.
Did Grok 4.7's price change from the previous version?
No. xAI's own pricing page lists Grok 4.7 at the same $2/$6 per million tokens as Grok 4.6, despite the new model using a larger 2.1-trillion-parameter base.
Are the benchmark comparisons between these models independently verified?
No. Every cross-company benchmark figure in launch announcements from Anthropic, OpenAI, and xAI is self-reported by the company that ran it, often against whichever competitor model was current at the time, not against each other's newest releases in a shared, independent test.
Why did three AI labs release new pricing within two days of each other?
Coverage of the launches frames it as intensifying price competition in AI model serving costs, with OpenAI's own GPT-6 Sol and Luna announcement arriving roughly 90 minutes after Anthropic's Opus 5.5 pricing went live, according to contemporaneous reporting.
Last verified: 14 September 2026. All three companies' pricing and model lineups are subject to frequent change — confirm current rates on each company's own pricing page before budgeting.