Short answer: the week of September 21–22, 2026 cut your API bill. OpenAI halved GPT-6 Sol to $2/$10 per million tokens and GPT-6 Luna to $0.10/$0.50. Anthropic shipped Claude Opus 5.5 at $4/$20 with $0.20 cache reads. xAI's Grok 4.7 held its price at $2/$6. OpenRouter's new Batch API takes about 50% off any batchable workload. On a 3,000-message month, the same work that cost $79 on GPT-5.6 Sol now costs $40.

Three model launches and one platform feature in about 36 hours. If you pay providers per token, and you haven't looked at your rate card since August, it's out of date. Here's what shipped, what each thing actually costs, and which numbers you should go change today.

What changed this week

Date (2026)	What	Who	Price move
Sep 21	Grok 4.7 launches	xAI	Holds $2/$6 per 1M tokens
Sep 22	GPT-6 Sol and Luna launch	OpenAI	Sol halves to $2/$10, Luna to $0.10/$0.50
Sep 22	Claude Opus 5.5 launches	Anthropic	$4/$20, down from Opus 5's $5/$25
Sep 22	Batch API ships	OpenRouter	~50% off per-token on 70+ models

A note on dates: OpenAI's announcement page carries no visible publication date, but it hit the Hacker News front page on September 22 and both Fortune and The New Stack covered it that day as landing within hours of Opus 5.5. Anthropic dates its own post September 22 and xAI dates Grok 4.7 September 21.

GPT-6 Sol and Luna: OpenAI halves its prices

The headline is the price cut, not the benchmarks. OpenAI says it's "reducing API prices for Sol and Luna by 50% compared with their GPT-5.6 promotional pricing," passing along caching and inference savings.

Per 1M tokens	Input	Output	Cache read	Cache write
GPT-6 Sol	$2.00	$10.00	$0.20	$2.50
GPT-5.6 Sol (old promo)	$4.00	$20.00	$0.40	$5.00
GPT-6 Luna	$0.10	$0.50	$0.01	$0.125
GPT-5.6 Luna (old promo)	$0.20	$1.20	$0.02	$0.25

Cache reads on GPT-6 are discounted 90%, which matters if your app reuses long prefixes. OpenAI also shipped a Prompt Caching Dashboard, a diagnostics tool for missed cache opportunities, explicit cache breakpoints, and (a quiet but real quality-of-life win) reasoning-effort and tool changes that no longer break your cache. GitHub says the caching work cut its fresh-processor token share by more than 50% across billions of requests.

Long-context pricing doubles input and adds half again to output ($4/$15 for Sol), so a 300K-token document job is not the same deal as chat traffic. Batch and Flex are half price again ($1/$5 for Sol); Fast mode doubles the base rate ($4/$20).

Availability details that matter for BYOK users: the API slugs are gpt-6-sol and gpt-6-luna, live now. ChatGPT Work and Codex got both models for Plus, Pro, Business, Enterprise, and Edu accounts; Free and Go accounts get Luna in the desktop app; the regular ChatGPT chat picker doesn't have them yet. And GPT-5.6 Sol's promotional pricing stays "available at least through November 21, 2026," so old slug names keep working at the old rate. If you want EU data residency on Sol or Luna, it's only available with Standard processing; regional endpoints carry a 10% uplift.

Claude Opus 5.5: flagship quality at 20% off list

Anthropic's new top model (page dated September 22) runs $4 input and $20 output per million tokens, 20% under Opus 5's $5/$25. The cache rates are where the win gets bigger: cache reads are $0.20 (5% of base input, against Opus 5's $0.50) and cache writes are $5 for a 5-minute cache or $8 for an hour. Output also generates more than 30% faster.

Per 1M tokens	Opus 5.5	Opus 5
Input	$4	$5
Output	$20	$25
Cache read	$0.20	$0.50
Cache write (5m)	$5	$6.25

Anthropic's claim is "40% less than Opus 5 on typical workloads," but read it carefully. On a fixed token count our math says 21% cheaper. The rest of that 40% comes from Opus 5.5 using fewer tokens per task on agentic work. If your workload is plain chat, expect the 21%. If it's long coding sessions, the token savings may deliver the rest.

Other things in the announcement worth knowing:

Fast mode (research preview, first-party API only) doubles output speed at $8/$40. Batch processing halves rates to $2/$10.
Subscription users on Pro, Max, Team, and seat-based Enterprise get higher five-hour usage limits plus one storable "rate limit reset."
Anti-distillation "preserved thinking" applies to API accounts created on or after August 31, 2026, and thinking mode can no longer be switched off. Zero data retention is still available, and EU AI Act watermarking applies.
The safeguards bite: most cybersecurity tasks get re-routed to Opus 4.8, and biology work needs the Life Sciences Verification Program. If you picked an Opus model specifically for security research, test this before you migrate.
Claude Sonnet 5.5 and Haiku 5.5 are promised "in the coming weeks."

Grok 4.7: same $2/$6, bigger model

xAI (now branding itself SpaceXAI) shipped Grok 4.7 on September 21 at exactly Grok 4.6's rates: $2 per million input tokens and $6 per million output, with a fast variant at twice the speed and twice the price. Context is 500K tokens and the knowledge cutoff is May 2026, per docs.x.ai.

The pitch is price-performance. xAI's own table puts Grok 4.7 at 46.3% on CursorBench 4.0 against Grok 4.6's 40.4%. The fine print came from elsewhere: VentureBeat and OpenTools both report that independent tests saw Grok 4.7 consume more tokens per task, so a "same price" model can still raise your bill. Flat rates are not flat costs.

OpenRouter Batch API: half price, and it works with your key

OpenRouter's Batch API (announced September 22) takes asynchronous workloads at roughly half the standard per-token price across 70+ models. You POST a list of requests to /api/v1/batches, poll until done, and collect results. Chat completions, responses, messages, and embeddings shapes are all accepted.

What the beta numbers look like: across 230k+ batches, the median finished in 7 minutes, 90% within an hour, 99% within 10.3 hours. Submissions between 5am and noon Pacific run slower (the slowest tenth took 2 to 4.5 hours); after 6pm Pacific the 90th percentile drops under 50 minutes. Each batch runs on one provider (cheapest eligible by default), results come back per-request so bad rows don't sink the job, and inputs plus results are kept 30 days.

The BYOK detail is the good one: with a provider key configured, batches route through your key and you pay only the BYOK fee. So the half-price discount layers onto your existing key setup instead of replacing it.

What it saves: a classification job of 100K rows at 1,500 input and 100 output tokens each on GPT-6 Sol costs $400 at standard rates and $200 through batch. If the rows can wait a few minutes, that's free money.

What the price war does to a real bill

Same workload everywhere: 3,000 messages a month, 4,000 input and 800 output tokens each, with 40% of input hitting cache and 5% written to cache (the same assumptions our cost breakdown for monthly ChatGPT API use uses).

Model	Per message	Per 1,000	Per month
GPT-6 Luna	$0.0007	$0.66	$1.98
Gemini 3.8 Flash*	$0.0049	$4.92	$14.76
Grok 4.7†	$0.0128	$12.80	$38.40
GPT-6 Sol	$0.0132	$13.22	$39.66
Claude Sonnet 5	$0.0132	$13.22	$39.66
Claude Opus 5.5	$0.0261	$26.12	$78.36
GPT-5.6 Sol (last week)	$0.0264	$26.44	$79.32

* Google's promo rate, which doubles January 1, 2027. Excludes $0.50 per million tokens per hour of cache storage. † Conservative: xAI's model table publishes no cache discount, so all input is priced at the base rate.

Three observations. First, GPT-6 Sol and Claude Sonnet 5 land on exactly the same $39.66, which makes the choice between them a quality question rather than a price question. Second, Opus 5.5 now costs what GPT-5.6 Sol cost last week. Third, the budget tier got absurd: Luna at $1.98 a month for that workload is cheaper than most people's coffee.

If you cache heavily, Opus 5.5's $0.20 reads change the math further, and the whole area rewards checking how a provider bills reads and writes before you pick.

Which model should you pick this week?

If you…	Use
Run high-volume simple tasks	GPT-6 Luna or deepseek-flash
Do serious coding on a budget	GPT-6 Sol or Claude Sonnet 5
Need top-tier agentic coding	Claude Opus 5.5
Live in Cursor and want one model	Grok 4.7 (watch the token counts)
Have a backlog that can wait	Any of the above through a batch API

The honest part

Vendor benchmark tables disagree with each other, as vendor tables do. xAI's launch table scores GPT-5.6 Sol at 72.7% on DeepSWE while OpenAI's own post reports GPT-6 Sol at 68.8% on what looks like the same eval family. Different harnesses, different effort settings, and no head-to-head anyone has published yet. Treat every score in this post as marketing until you run your own evals.

Also, prices move fast right now. Gemini's promo rates double in January, GPT-5.6 Sol's promo expires (at least) November 21, and Sonnet 5.5 lands within weeks. Any comparison table published this month has a shelf life.

And if you use one provider, one model, and modest volume, none of this requires action. Your provider's dashboard already shows your spend, and switching models for a 21% cut on $15 a month isn't worth an afternoon.

FAQ

When do GPT-6 Sol and Luna appear in ChatGPT?

Not in the regular chat picker yet. They're in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu accounts as of September 22, with a gradual rollout through the day. Free and Go users get Luna in the desktop app. API access via gpt-6-sol and gpt-6-luna is live now.

Is Claude Opus 5.5 really 40% cheaper than Opus 5?

On agentic workloads where it uses fewer tokens per task, maybe. On identical token counts the cut is 21%, and cache-heavy work does better still because reads dropped from $0.50 to $0.20 per million.

Does OpenRouter's Batch API work with my own provider key?

Yes. With a provider key configured, batches route through it and you pay the BYOK fee instead of full price. Each batch runs on a single provider, and the discount varies by model.

Should I switch everything to GPT-6 Luna?

Only if your tasks tolerate it. Luna is a speed-and-cost model. The sensible move this month is routing: Luna-class models for volume work, Sol or Sonnet 5 for real drafts, Opus 5.5 for the hard 10%. Our guide to chatting with multiple models covers that workflow.

Did anything change for DeepSeek or Gemini this week?

No launches from either. DeepSeek's deepseek-flash (V4.1-Flash) still runs $0.15/$0.60 off-peak and Gemini 3.8 Flash still runs $0.75/$3.75 through December 31. Both look in the mirror at these new prices, though.

If you want to try the new models without juggling four dashboards, Multimodel Chat connects OpenAI, Anthropic, Google, xAI, and OpenRouter keys in one workspace. How it works →

Start your free trial → — 7 days, all providers, no credit card required.

New AI Model Release Roundup (2026): Sol, Opus 5.5, Grok 4.7