Same-Day Price Cuts at the Frontier: Claude Opus 5.5 Drops to $4/$20 as GPT-6 Sol Lands at $2/$10

September 22, 2026 was a strange Tuesday for the AI business. Within hours of each other, Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, while OpenAI released GPT-6 Sol at $2/$10 and GPT-6 Luna at $0.10/$0.50¹²³. The two launches landed ten days after Anthropic CEO Dario Amodei published a blog post urging the industry to slow the pace of new feature releases, a call OpenAI CEO Sam Altman publicly backed on X⁴.

What slowed down was the capability race. The price race sped up.

Three weeks that rewrote the price list

The September calendar tells the story better than any single announcement. OpenAI opened the month with a flagship, SpaceXAI (formerly xAI) slipped in a mid-tier entry, and then both labs cut on the same day.

Table 1: Timeline of the September 2026 price war

Date	Event
2026-09-03	OpenAI releases flagship GPT-6 Astra at $10 input / $50 output per million tokens56
2026-09-12	Amodei urges slower releases in a blog post; Altman backs him on X4
2026-09-21	SpaceXAI ships Grok 4.7 at $2/$6 for requests under 200K tokens7
2026-09-22	Anthropic ships Claude Opus 5.5 at $4/$20; OpenAI ships GPT-6 Sol ($2/$10) and GPT-6 Luna ($0.10/$0.50)123

Opus 5.5 comes in 20% below Opus 5, which listed at $5/$25¹⁸. Anthropic says a typical workload should cost about 40% less than on Opus 5, combining the 20% price cut with fewer tokens per task, and claims output speed is more than 30% faster¹². Cache reads drop to $0.20 per million tokens (5% of input price; Opus 5 charged $0.50 at 10%, a 60% cut)¹⁸. Batch mode runs $2/$10, and a Fast mode at $8/$40 is first-party API only¹. Sonnet 5.5 and Haiku 5.5 are due in the coming weeks².

OpenAI’s counter landed the same day. GPT-6 Sol targets complex professional, coding, and agentic work; GPT-6 Luna targets high-volume, low-cost jobs like extraction and summarization³. Both are priced 50% below the GPT-5.6 promo rates (Sol’s promo was $4/$20) and ship in ChatGPT Work, Codex, and the API, not in consumer Chat³⁹. OpenAI says Sol’s error rate is roughly half its predecessor’s, that Luna at high reasoning reaches GPT-5.6 Sol’s level at about 1% of the cost, and that prompt caching can cut cached-prompt cost by up to 90%³. The company attributes the reductions to “cache and inference improvements,” with savings passed back to users⁴.

The rate card, side by side

Table 2: API rate card, USD per million tokens ⁶⁷¹⁸¹⁰

Model	Input	Output	Cache read	Released
Claude Opus 5.5	$4	$20	$0.20 (5%)	2026-09-22
Claude Opus 5	$5	$25	$0.50	2026-07
Claude Fable 5.1	$10	$50	$0.25	—
GPT-6 Astra	$10	$50	$1.00 (10%)	2026-09-03
GPT-6 Sol	$2	$10	—	2026-09-22
GPT-6 Luna	$0.10	$0.50	—	2026-09-22
GPT-5.6 Sol (promo)	$4	$20	$0.40	promo through at least 2026-11-21
Grok 4.7 (≤200K)	$2	$6	$0.50	2026-09-21

Against Astra, Opus 5.5 lists at 40% of the flagship’s input and output rates (a 2.5x gap both ways), and its cache read price is one fifth of Astra’s, $0.20 versus $1.00¹¹. OpenAI has not published Sol or Luna cache pricing, so those cells stay blank.

Why the price war started now

Three pressures converged this month.

The first is external. Chinese open-weight models, which customers can run and customize locally, keep taking share. Vercel’s usage data shows these models winning market share away from US labs, with Alibaba, Moonshot AI, and DeepSeek named as the pressure sources¹²¹³. OpenAI has already cut prices repeatedly in recent months¹³.

The second is internal, and different for each lab. Anthropic is preparing an autumn IPO that is expected to be the largest tech IPO ever, and market speculation ties the model launch directly to the S-1 catalyst¹³. OpenAI saw growth slacken earlier this year and says it is unlikely to go public before 2027; price cuts are its lever to rekindle growth¹³.

The third is structural. Ramp chief economist Ara Kharazian put it bluntly: “OpenAI and Anthropic are engaged in a price war, which drives AI prices down and reduces their ability to profit from and raise model prices.” He describes a war on two fronts: cheaper models like Opus 5.5 and GPT-6 Sol/Luna on one side, direct flagship price cuts on the other⁴. (His comments reach English readers through translation, so treat the wording as reported commentary.) 财联社’s read of the same-day launches is that competition is shifting from a flagship capability arms race to a combined performance-and-price contest for cost-sensitive enterprise customers³.

Then there is the timing. Ten days after both CEOs publicly urged slowing frontier progress, both labs dropped price-cutting models on the same day⁴¹³. Whatever “slow down” meant in public, the contest moved from capability to cost.

The sticker price is not the bill

For agent-era workloads, the input and output columns in Table 2 matter less than two other things: cache rates and long-context rules.

Agents re-read the same context every turn, so cache reads dominate the bill: Opus 5.5 charges 5% of input price, Astra 10%¹¹. Above 272K input tokens, Astra reprices the entire request (2x input and cache, 1.5x output, or $20/$75 for that call); Opus 5.5 bills its full 1M-token window flat, and Grok 4.7 jumps the whole request to $4/$12 past 200K¹¹¹⁴.

Digital Applied’s worked example uses identical token counts (8M cache reads, 400K uncached input, 600K cache writes, 300K output): Opus 5.5 totals $12.20, Astra under the cliff $34.50, Astra over the cliff $61.50, roughly a 5x spread once the cliff trips¹¹. For one agent turn (100K input with 80K cached, 10K output), alphacorp.ai puts Opus 5.5 at about $0.30 versus Astra’s $0.78, a 62% saving; at 10,000 turns a day, roughly $2,960 versus $7,800¹⁵.

Cheaper per token, more expensive per task

Here is the trap. Artificial Analysis (v4.3.2) ranks Opus 5.5 at maximum effort first of 212 models on its Intelligence Index, 58 points against Astra’s 53 and Grok 4.7’s 46¹⁴¹⁶. But running the full index, Opus 5.5 burned about 260 million output tokens costing $8,708; Astra used 60 million costing $5,324. Per task, that is roughly 119K output tokens against 27K, about 4x more¹⁴¹⁶.

Do the arithmetic and the 60% sticker discount shrinks fast: at maximum effort, Opus 5.5’s per-task output cost can actually exceed Astra’s, about $2.38 versus $1.49 for output tokens alone¹⁴¹⁶. Anthropic’s cost claims, including beating Astra on FrontierCode at roughly one fifth of the cost per task and matching it on Terminal-Bench 4.0 at about 40% of the cost, are measured at the default medium effort tier, and they come from the winner’s own launch material rather than independent reproduction⁸¹¹. OpenAI’s competing claim, that GPT-6 Sol scores 6.3% above Claude Opus 5 on AutomationBench at 9% of the per-task cost, is likewise vendor math³.

The practical rule: compare models on cost per completed task at the effort tier you will actually run, and treat list prices as a starting hint.

The benchmarks are noisier than the price gap

The price tables are exact. The benchmark tables are anything but.

On Terminal-Bench 4.0, Anthropic’s run shows Opus 5.5 at 66.4% (xhigh effort) against Astra’s 57.9% (high effort, OpenAI’s own figure). When Artificial Analysis ran both models independently at matched settings, the result was a dead tie: 59.6 versus 59.6¹⁶¹¹. Vendors also disagree when scoring the same model. On FrontierCode Main, Anthropic scores Opus 5 at 48.0% while OpenAI scores it at 53.4%, a 5.4-point gap, nearly five times Opus 5.5’s 1.1-point lead over Astra there. OpenAI footnoted that Astra ran with a Codex-style developer message¹¹.

The genuine leads match each model’s positioning. Astra holds reasoning and science turf: GPQA Diamond at 96.0%, FrontierMath Tier 4 v2 at 97.6%, ARC-AGI-3 at 62.71% (ARC Prize verified), Terminal-Bench-Science at 64.6% versus 58.7%, with AutomationBench essentially level at 41.4% versus 40.0%, inside the noise band¹⁵. Opus 5.5’s leads beyond noise sit elsewhere: Terminal-Bench 4.0 by 8.5 points, GDPval-AA at 1846 versus 1542 Elo (a 304-point gap), and HLE with tools at 67.7% versus 57.2%¹¹. Standard error on these agentic benchmarks runs about ±3.5 to 5 points, so anything smaller than that is a coin flip¹⁵.

What this means for developers

If you run agents at volume, this month genuinely changed your bill. Three weeks ago, none of the new options existed: no $2/$10 mid-tier, no $0.10/$0.50 light tier, no top-tier coding model at $4/$20 with 5% cache pricing⁶¹³.

Before switching anything, keep three habits. First, re-run your own tasks at matched effort tiers on both vendors; the benchmark gap you read about is probably smaller than run-to-run noise. Second, model your cache traffic: if your workload re-reads large contexts, cache rates and the long-context cliff decide your bill before sticker prices do. Third, measure cost per completed task, including retries and error rates, rather than cost per token. OpenAI’s claim that Sol halves its predecessor’s error rate matters to your invoice only if it holds on your tasks³.

References

Conclusion

Two labs, one Tuesday, and a frontier price list rewritten in a day. Anthropic cut its top model by a fifth and slashed cache rates; OpenAI answered with a mid-tier at half the previous promo price and a light tier that costs almost nothing. Both moves landed days after their CEOs asked the industry to slow down, which says where the competition actually moved.

The sticker cuts are real, and so are the traps. Cache rules, context cliffs, and effort tiers decide the invoice as much as the rate card, and the benchmark gap between these models is smaller than the noise in how they get measured. The labs have made their opening bids; the next move will probably be about price too.

Footnotes

1. Anthropic Model Docs — Claude Opus 5.5 overview: $4/$20 pricing, $0.20 cache reads, Batch and Fast mode rates, workload cost and speed claims https://platform.claude.com/docs/en/models/opus-5-5/overview ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

2. NBD — Coverage of the Claude Opus 5.5 and GPT-6 Sol/Luna same-day launches; Sonnet 5.5 and Haiku 5.5 due in coming weeks (in Chinese) https://www.nbd.com.cn/articles/2026-09-23/4589497.html ↩ ↩² ↩³ ↩⁴

3. Eastmoney / 财联社 — GPT-6 Sol and Luna launch details: 50% below GPT-5.6 promo pricing, error-rate and Luna reasoning claims, up-to-90% prompt caching savings, and the shift to performance-and-price competition (in Chinese) https://finance.eastmoney.com/a/202609233881630823.html ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

4. Gelonghui — Report on Amodei’s slow-down blog post, Altman’s backing on X, Ramp economist Ara Kharazian’s price-war commentary, and OpenAI’s “cache and inference improvements” statement (in Chinese, via translation) https://m.gelonghui.com/p/6780775 ↩ ↩² ↩³ ↩⁴ ↩⁵

5. DataNorth — OpenAI launches GPT-6 Astra at $10/$50 per million tokens, September 3, 2026 https://datanorth.ai/news/openai-launches-gpt-6-astra ↩

6. OpenAI API Pricing — Official per-token pricing for GPT-6 Astra, GPT-6 Sol, and GPT-6 Luna https://developers.openai.com/api/docs/pricing ↩ ↩² ↩³

7. TECHSY — GPT-6 Astra vs Claude Opus 5.5 vs Grok 4.7 comparison, including Grok 4.7’s $2/$6 pricing under 200K tokens and its 200K price jump https://techsy.io/en/blog/gpt-6-astra-vs-opus-5-5-vs-grok-4-7 ↩ ↩²

8. Kingy AI — Claude Opus 5.5 specs, benchmarks, and pricing comparison, including Anthropic’s FrontierCode and Terminal-Bench cost-per-task claims https://kingy.ai/blog/claude-opus-5-5-specs-benchmarks-pricing-comparison/ ↩ ↩² ↩³ ↩⁴

9. OpenAI ChatGPT Rate Card — Enterprise token-based pricing and product availability for GPT-6 Sol and Luna in ChatGPT Work and Codex https://help.openai.com/en/articles/20001415-chatgpt-rate-card-enterprise-token-based-pricing ↩

10. Anthropic Pricing — Official Claude model pricing, including Opus 5, Opus 5.5, and Fable 5.1 rates https://platform.claude.com/docs/en/about-claude/pricing ↩

11. Digital Applied — Claude Opus 5.5 vs GPT-6 Astra comparison: 40% rate gap, cache pricing, the 272K long-context cliff, worked cost examples, vendor benchmark disagreements, and per-model leads https://www.digitalapplied.com/blog/claude-opus-5-5-vs-gpt-6-astra-comparison ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

12. Gate News — Report on Claude Opus 5.5’s 40% cost reduction, Anthropic’s enterprise-cost framing, and Vercel data on Chinese open-weight models taking share (in Chinese) https://www.gate.com/zh/news/detail/OPENAI/anthropic-releases-claude-opus-55-with-costs-down-40-24489138 ↩

13. Caijing CN (cjcn) — Analysis of Anthropic’s autumn IPO preparations, OpenAI’s growth slowdown and pre-2027 IPO stance, repeated OpenAI price cuts, and open-weight pressure from Alibaba, Moonshot AI, and DeepSeek (in Chinese) https://www.cjcn.com.cn/news/show-491793.html ↩ ↩² ↩³ ↩⁴ ↩⁵

14. Codersera — Opus 5.5 vs GPT-6 Astra vs Grok 4.7 comparison covering Artificial Analysis Intelligence Index scores, output-token burn, and per-task cost at maximum effort https://codersera.com/blog/claude-opus-5-5-vs-gpt-6-astra-vs-grok-4-7-2026/ ↩ ↩² ↩³ ↩⁴

15. Alphacorp — Opus 5.5 vs Astra benchmarks and pricing: single-agent-turn cost math, daily volume totals, and Astra’s reasoning and science benchmark leads https://alphacorp.ai/blog/claude-opus-5-5-vs-gpt-6-astra-benchmarks-pricing-and-which-is-better ↩ ↩² ↩³

16. DataStudios — Full comparison report on pricing, benchmarks, and effort levels: Artificial Analysis token usage, per-task output costs, and the independent Terminal-Bench 4.0 tie https://www.datastudios.org/post/claude-opus-5-5-vs-gpt-6-astra-complete-comparison-and-report-on-pricing-benchmarks-effort-levels ↩ ↩² ↩³ ↩⁴