Opus 5.5 Reset the API Prices. Sol Halved Them.

On September 22, Anthropic and OpenAI both cut API prices. CNBC framed it as a response to cheaper open-weight models. The numbers that matter are narrower than that frame. Anthropic’s Opus 5.5 page puts input at $4 per million tokens and output at $20, down from $5 and $25 on Opus 5. Engadget puts GPT-6 Sol at $2 and $10, and GPT-6 Luna at $0.10 and $0.50. Those are the stickers. The bill is still mostly cache.

We already looked at Grok 4.7’s price list versus the bill. That post was one vendor keeping a sticker and still charging you for how the model behaves. This one is two vendors changing the sticker on the same Tuesday, and still leaving the expensive lines in the footnotes. The August comparison of Opus 5 and GPT-5.6 is a different generation. Do not reuse it as a shopping list.

What the API prices actually say

Anthropic’s table, per million tokens:

Line	Opus 5.5	Opus 5
Input	$4	$5
Output	$20	$25
Cache reads	$0.20	$0.50
Cache writes	$5	$6.25

Input and output are 20% cheaper. Cache reads are 60% cheaper. Anthropic says cache reads are the majority of the cost on agentic and coding work, and that their tests at default settings land about 40% below Opus 5 on typical workloads. The 40% is not the 20%. It is the 20%, plus fewer tokens per task, plus that cache-read cut. If your traffic is one-shot prompts with no cache, you do not get the 40%. You get the sticker.

Fast mode is a separate product. Up to 2.5x speed, $8 per million input and $40 per million output. That is double the standard Opus 5.5 rates. Anyone who turns it on and then quotes the $4/$20 line is describing a model they are not running.

OpenAI’s cut is louder and applies to different rungs. CNBC says Sol and Luna are 50% below GPT-5.6 promotional API prices. Sol sits under Astra and is aimed at coding and other complex work. Luna is the high-volume tier, for extraction and summarization. Coursiv, citing OpenAI’s pricing page on September 22, publishes the comparison that makes the “50%” uneven:

Line	GPT-6 Sol	GPT-5.6 Sol promo	GPT-6 Luna	GPT-5.6 Luna
Input	$2.00	$4.00	$0.10	$0.20
Output	$10.00	$20.00	$0.50	$1.20
Cached input	$0.20	$0.40	$0.01	$0.02
Cache writes	$2.50	$5.00	$0.125	$0.25

Luna’s output did not fall by half. It fell from $1.20 to $0.50, which is closer to a 58% cut. Sol’s lines did halve, including cache. Long context, over 272,000 input tokens, is billed higher: Coursiv lists Sol at $4 and $15, Luna at $0.20 and $0.75, against promo rates of $8/$30 and $0.40/$1.50. Fast mode is 2x. Batch and Flex are 50% of standard. Regional data-residency processing adds 10%. If your “50% off” spreadsheet ignores the long-context bump, it is a press-release spreadsheet.

Felloai quotes OpenAI’s September 22 post: the company said better caching and inference paid for the cut, and that the 50% was versus GPT-5.6 promotional pricing, not versus some imaginary list price from 2024. Promotional is the word to keep. A promo-to-standard comparison can look like a sale when it is a new rung replacing a discount.

Astra did not join the sale. Coursiv says GPT-6 Astra stays at $10 input and $50 output per million. Felloai dates Astra’s launch to September 3, as the single GPT-6 model, with Sol and Luna arriving on the 22nd as the rest of the ladder. Same family, three prices, and only two of them moved.

The benchmark table is not a receipt

Anthropic published a comparison that includes its own models and OpenAI’s. I am treating it as Anthropic’s table, not as a neutral lab.

On Terminal-Bench 4.0, Opus 5.5 scores 66.4%. Fable 5.1 is 55.8%. Opus 5 is 52.3%. GPT-6 Astra is 57.9%. GPT-5.6 Sol is 37.3%. On FrontierCode v1.1 (Main), the order is 54.4, 50.3, 48.0, 53.3, 47.5. Astra is close. On CursorBench 4.0, Opus 5.5 is 57.8% and GPT-5.6 Sol is 41.7%. Astra has no figure in that cell.

GDPval-AA v2.1, their knowledge-work Elo across 44 occupations: Opus 5.5 at 1846, Fable 5.1 at 1735, Opus 5 at 1708, GPT-5.6 Sol at 1588, Astra at 1542. AutomationBench is the awkward row. Astra leads at 41.4%. Opus 5.5 is 40.0%. Everyone else is in the high 20s or low 30s. A model can be the coding pick and still lose a workflow bench by a point.

Anthropic’s own cost claims, which Engadget also relays for the coding benches: at default effort on FrontierCode, Opus 5.5 beats Astra at roughly 20% of the cost per task. On Terminal-Bench 4.0 it matches Astra for about 40% of the cost. On CursorBench it beats GPT-5.6 Sol by 11 points for about a third of the cost. On GDPval, medium effort beats Astra at max effort for about a fifth of the cost per task.

Read the effort clause. “Beats Astra at max effort while we used medium” is a real claim and a slippery one. If your production setting is max effort, or fast mode, you are not in that sentence. Anthropic also says the gap between Opus 5.5 and Fable 5.1 is narrower in their own use than the scores suggest, and that benchmark margins at this level are a less reliable guide to real-world differences. I would keep that sentence next to the 66.4%. The vendor who published the win also published the caveat.

OpenAI’s quality claims, via Engadget, are thinner. Sol makes about half as many mistakes as its predecessor. Sol matches Fable 5.1 at a lower cost. “Half as many mistakes” without the test name is a slogan until someone posts the set. Anthropic at least named the benches.

There is a science footnote on Anthropic’s page that should humble both tables. On Terminal-Bench-Science 0.1, the public leaderboard has Claude Opus 5 at 30.0%. Anthropic’s setup reproduces it at 29.0%, inside a standard error of about 3.5 to 5 points. The Astra figure is “as reported by OpenAI.” A 1-point gap is noise. A 10-point gap on Terminal-Bench 4.0 is not the same kind of number, but the footnote is a reminder that harness and trial count move the score.

Cache, context, and the lines people skip

Both labs are selling the cache discount harder than the sticker. Anthropic’s cache read fell from $0.50 to $0.20. OpenAI’s Sol cached input is $0.20 against a $2 input price, which is the 90% read discount Felloai describes, or 10% of the input rate in Coursiv’s wording. Same arithmetic. Cache writes did not get that treatment. Opus 5.5 cache writes are $5, a quarter more than the $4 input price. Sol cache writes are $2.50 against $2 input. Writing the cache is not the cheap part. Reading it is.

That matters if your agent rewrites context every turn instead of reusing it. The cheap line never fires. You pay input, or you pay a cache write, and the blog post you read on September 22 does not describe your invoice. We have been through this shape of disappointment with Grok. The mechanism is not unique to xAI.

Context windows are not the discount either. Felloai, citing OpenAI’s API docs, gives both Sol and Luna a 1,050,000-token window and a 128,000-token output cap. Knowledge cutoffs differ: April 20, 2026 for Sol, May 18, 2026 for Luna. Reasoning effort runs from none through low, medium (the default), high, xhigh, and max. Coursiv says built-in tools and function calling go to the Responses API, and that a no-reasoning setting is the condition for some of that path. If you need tools, you may not be on the cheapest effort. Price the configuration you will actually call.

Long context is the other trap. Cross 272,000 input tokens on Sol and Coursiv’s table moves you from $2/$10 to $4/$15. That is still below Astra’s $10/$50, and it is no longer the number in the launch tweet. A coding agent that pastes a repo into the prompt can leave the headline rate without anyone changing the model string.

Where you can even call them

API availability and chat availability split on day one. Coursiv lists the API model names as gpt-6-sol and gpt-6-luna. Felloai says they are in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu, and that OpenAI said they are not yet in ordinary Chat. Free and Go users can reach Luna in the desktop app. Nobody in these write-ups has a date for the normal chat window. Felloai also notes that a GPT-6 Terra was not announced. GPT-5.6 still has a Terra rung. GPT-6, so far, does not. If your router expects three names and the middle one 404s, that is why.

Inside ChatGPT Work, Coursiv lists a credit price that is not the API price: Sol at 50 credits per million input tokens and 250 per million output, Luna at 2.5 and 12.5. A team that “switched to Sol” in the Work UI is not necessarily paying $2 and $10. They may be spending credits against a seat. Those are different budgets. Mix them in one spreadsheet and the 50% figure becomes fiction.

Opus 5.5 has its own non-price changes, and they will break integrations that treated thinking as optional. Anthropic says preserved thinking, the control they introduced with Fable 5.1, applies to Opus 5.5 for API accounts created on or after August 31, 2026. It stops clients from editing prior context to pull reasoning out. Thinking mode can no longer be switched off. Zero data retention is still available, as on earlier Opus models. Watermarking is there for the EU AI Act, same posture as Fable 5.1. If your pipeline stripped thinking blocks or assumed a no-thinking mode, the price cut is the small part of the migration. The contract change is the part that fails the build.

Subscription plans get a separate sweetener: higher five-hour usage limits on Pro, Max, Team, and seat-based Enterprise, plus a rate-limit reset users can save and spend when they want. That is not an API discount. It is a consumer and seat-plan change. Do not quote it in a procurement memo about token rates.

What I would switch, and what I would leave

I would move a high-volume extract job to Luna if the task is shaped, repeated, and tolerant of a model OpenAI itself describes as the volume tier. At $0.10 and $0.50, with a million-token window, the old reason to keep a mid-tier model on classification was the price. That reason got worse for the mid-tier on September 22. I would still sample outputs. “Half as many mistakes” is not a test I can rerun from the Engadget sentence.

I would price Sol against Opus 5.5 on a coding agent only after I log cache-read share and effort level for a week. Sol at $2/$10 versus Opus 5.5 at $4/$20 looks like a clean half. It is not, once Opus is hitting a $0.20 cache read on most of the prompt and Sol is paying $2.50 to write cache or $4 because the prompt crossed 272k. Anthropic’s “20% of Astra’s cost per task” is a task cost, not a token cost. Task cost is the number a finance person should ask for. Token cost is what the launch posts led with.

I would not move a workload to Astra to “get the flagship” unless AutomationBench-shaped work is the job. On that one row Astra still leads. On Terminal-Bench 4.0 and FrontierCode, Anthropic’s table has Opus 5.5 ahead, and Astra’s sticker did not fall. Paying $10 and $50 for a model that lost the coding rows in a competitor’s table is a choice. It is not the default the September 22 posts imply.

Open-weight models are the pressure CNBC named, and we have covered that split before in the two-track race. The September 22 cuts are what the paid side does when the free side is good enough for the boring calls. They do not make the paid invoice simple. They add rungs, a long-context cliff, a fast-mode multiplier, and a cache line that only helps if your client actually reuses context.

If you change one thing after reading the posts, change the dashboard. Split standard tokens, cache reads, cache writes, and long-context tokens before you celebrate a percent. The percent in the headline is real. It is just not the percent on the line that dominates an agent bill.