AI Tools
What does OpenAI API pricing look like after the GPT-6.1 Sol launch?
OpenAI API pricing now runs on three flagship rates: GPT-6.1 Sol costs $2.00 per 1M input tokens and $10.00 output, GPT-6 Astra costs $10.00 and $50.00, and GPT-6 Luna costs $0.10 and $0.50, per OpenAI's pricing page checked September 30, 2026. Claude's flagship Fable 5.1 charges the same $10 and $50 as Astra.
What matters
- GPT-6.1 Sol is $2.00 per 1M input tokens and $10.00 per 1M output tokens, exactly one-fifth of GPT-6 Astra's $10.00 and $50.00 at the Standard tier and short context.
- Claude Sonnet 5.5 lists the same $2.00 input / $10.00 output as Sol, and Claude's flagship Fable 5.1 lists the same $10.00 / $50.00 as Astra, so OpenAI's new tier undercuts Anthropic's top model five times over.
- Cached input is 20x cheaper than fresh input on Sol ($0.10 vs $2.00 per 1M), but a cache write costs $2.50, a 25% premium over sending the same text as plain input. Every OpenAI and Anthropic flagship carries that same 25% write premium.
- Crossing into the long-context band doubles every OpenAI input-side rate and raises output by 50%: Sol moves from $2.00 / $10.00 to $4.00 / $15.00.
- A month of 50M input tokens and 5M output tokens costs about $150 on Sol, $750 on Astra and $7.50 on Luna at Standard rates, and the Batch and Flex tiers halve those numbers.
OpenAI API pricing in one table (checked September 30, 2026)
OpenAI API pricing is billed per million tokens (MTok) and split into four line items per model: input, cached input, cache writes and output. The flagship line has three tiers after the September 29 launch: GPT-6 Astra at the top, GPT-6.1 Sol one rung below, and GPT-6 Luna as the cheap volume model. All three are reasoning models with a 1.05M-token context window, 128K maximum output and an April 30, 2026 knowledge cutoff, per OpenAI's model pages.
Every OpenAI flagship rate comes in two bands, short context and long context. OpenAI's pricing page does not state where the short band ends, so check the model page for your own prompt sizes before you budget at the lower number. Anthropic publishes one rate per model with no band split, which is why the Claude rows below carry no band label.
GPT-6.1 Sol vs GPT-6 Astra: what the fifth-of-the-price claim means
The launch framing was about savings, and the arithmetic backs it: Astra's $10.00 input rate is exactly five times Sol's $2.00, and its $50.00 output rate is exactly five times Sol's $10.00. The discount does not shrink on the output side, which is where most agent and coding workloads spend their tokens.
The two models also share the plumbing. Same 1.05M context window, same 128K max output, same April 30, 2026 knowledge cutoff, and the same tool list (functions, web search, file search, computer use) on OpenAI's model pages. Moving a workload from Astra to Sol is a model ID change, not a re-architecture.
What you give up is the accuracy ceiling. OpenAI describes Sol as "near-Astra performance for complex work at a lower cost" and Astra as "our most capable model for the most demanding work". Near is not the same as equal. Treat Sol as the default and Astra as the escape hatch for work that actually fails on Sol, not the other way around.
OpenAI's own price history makes the launch look less like a discount and more like a reset. The same page still lists GPT-5.6 Sol at $4.00 input and $20.00 output per 1M tokens and notes that its promotional pricing is available at least through November 21, 2026. GPT-6.1 Sol is half of that on both sides, so an upgrade does not need an accuracy win to pay for itself.
GPT-6.1 Sol vs Claude Sonnet 5.5 and Fable 5.1
Anthropic's pricing page makes the more interesting comparison. Claude's flagship Fable 5.1 is $10.00 input and $50.00 output per MTok, an exact match for GPT-6 Astra, so GPT-6.1 Sol undercuts Anthropic's top model by five times. The closer fight is one tier down, where Claude Sonnet 5.5 lists $2.00 input and $10.00 output, identical to Sol down to the cent.
Two structural differences decide that matchup. Anthropic publishes a single input and output rate per model with no short/long split, which makes it easier to forecast, while OpenAI's input-side rate jumps once a request counts as long context. On caching, Anthropic's reads are cheaper at the top end: $0.25 per MTok on Fable 5.1 and $0.20 on Sonnet 5.5, against $1.00 on Astra and $0.10 on Sol.
The one thing both vendors agree on is that writing a cache costs more than sending the same text as input, at exactly the same 25% premium: $12.50 against $10.00 input on Fable 5.1, $2.50 against $2.00 on Sonnet 5.5, $2.50 against $2.00 on Sol. Compliance adds roughly the same surcharge on both sides too. Anthropic charges 1.1x for US-only inference, and OpenAI charges a 10% uplift on regional data-residency endpoints for models released on or after March 5, 2026, plus 10% on FedRAMP endpoints.
Cache writes, cached reads and long context decide your bill
Two numbers in that table matter more than their size suggests. On Sol, cached input is $0.10 per 1M tokens against a $2.00 input rate, a 95% discount, while a cache write costs $2.50, which is 25% more than sending the same tokens as ordinary input. The read discount is 90% on Astra and Luna, 95% on Sol.
The practical rule: caching is a discount only when the same prefix comes back. An agent harness that resends a long system prompt and repository context on every turn collects the $0.10 reads and wins big. A one-shot script that writes a cache and never reads it back pays $0.50 per million tokens over the plain input rate for nothing.
Long context is not one multiplier
OpenAI does not apply a single multiplier when a request moves into the long-context band. Input, cached input and cache writes all double, while output rises 50%.
- GPT-6.1 Sol: $2.00 / $10.00 short context becomes $4.00 / $15.00 long context.
- GPT-6 Astra: $10.00 / $50.00 becomes $20.00 / $75.00.
- GPT-6 Luna: $0.10 / $0.50 becomes $0.20 / $0.75.
For a workload that reads a lot and writes a little, that is expensive. A month of 50M input tokens and 5M output tokens costs $150 on Sol at short-context rates and $275 at long-context rates, an 83% jump that comes from crossing a band boundary rather than from doing more work.
Batch, Flex and Fast: three tiers that move the price
OpenAI sells the same models at several service levels, and the pricing page's tabs are easy to miss. Batch, described as being for large numbers of requests that are not time-sensitive, lists exactly half the Standard rates: Sol at $1.00 input and $5.00 output, Astra at $5.00 and $25.00, Luna at $0.05 and $0.25 per 1M tokens.
Flex lists the same halved rates and is described as lower prices with higher latency, so the real choice between Batch and Flex is about request shape, not price. Fast sits above Standard for priority processing and Ultrafast sits above Fast; on the current page Ultrafast lists only GPT-6 Astra, at $60.00 input and $300.00 output per 1M tokens in short context.
If any part of your pipeline is a nightly evaluation sweep, a backfill or a batch summarisation job, moving it to Batch is a straight 50% cut with no model change and no prompt change.
Who should switch to GPT-6.1 Sol, and who should stay on Astra
Switch when token spend is the constraint and the work is repetitive: agentic coding, document extraction, RAG answering, batch classification. A team burning $750 a month on Astra for a 50M-input workload pays $150 on Sol for the same token volume, and routing the genuinely simple traffic to Luna drops it to $7.50.
Stay on Astra when the workload leans on something Astra alone offers, when your evaluation set was calibrated on Astra, or when a single failure costs more than the token bill. Ultrafast mode also lists Astra only, so latency-critical pipelines have no Sol option yet. If you are already standardised on Claude, the equivalent move is comparing Sonnet 5.5 against Fable 5.1 on the same terms, since the two vendors now price their mid-tier models identically.
One caution about the numbers above: they are list rates, not your bill. OpenAI bills web search separately at $10.00 per 1,000 calls plus the tokens those results consume, and ChatKit file and image upload storage at $0.10 per GB-day after the first free gigabyte per account per month. Agent workloads can carry more cost in tool calls than in tokens.
How to estimate your own OpenAI API bill before switching
Add four products, then divide by a million: input tokens times the input rate, cached reads times the cached rate, cache writes times the write rate, and output tokens times the output rate. Pull the token counts from a full month of production traffic rather than a sample, because cache hit rates vary too much between workloads to guess.
Two shapes are worth checking before you commit. Input-heavy pipelines get the most from Sol's $0.10 cached reads. Output-heavy pipelines get the most from the fact that Sol's output discount is a full five times, not a partial one. If your traffic is mostly short classification or extraction calls, price it on Luna at $0.10 input / $0.50 output before assuming Sol is the cheap option.
At a glance
| Model | Input | Cached input | Cache writes | Output |
|---|---|---|---|---|
| GPT-6.1 Sol, short context | $2.00 | $0.10 | $2.50 | $10.00 |
| GPT-6.1 Sol, long context | $4.00 | $0.20 | $5.00 | $15.00 |
| GPT-6 Astra, short context | $10.00 | $1.00 | $12.50 | $50.00 |
| GPT-6 Astra, long context | $20.00 | $2.00 | $25.00 | $75.00 |
| GPT-6 Luna, short context | $0.10 | $0.01 | $0.125 | $0.50 |
| GPT-6 Luna, long context | $0.20 | $0.02 | $0.25 | $0.75 |
| Claude Fable 5.1 | $10.00 | $0.25 | $12.50 | $50.00 |
| Claude Sonnet 5.5 | $2.00 | $0.20 | $2.50 | $10.00 |
| Claude Haiku 4.5 | $1.00 | $0.10 | $1.25 | $5.00 |
FAQ
Is GPT-6.1 Sol cheaper than GPT-6 Astra on the OpenAI API?
Yes, five times cheaper on list rates: $2.00 against $10.00 per 1M input tokens, and $10.00 against $50.00 per 1M output tokens at the Standard tier and short context.
Does OpenAI charge more for long context?
Yes. In the long-context band OpenAI doubles the input-side rates (input, cached input and cache writes) and raises output by 50%. GPT-6.1 Sol goes from $2.00 input and $10.00 output to $4.00 and $15.00 per 1M tokens.
Can I pay less than list price for the OpenAI API?
Yes, two ways on the same page: the Batch tier halves token rates for requests that are not time-sensitive (Sol at $1.00 input / $5.00 output), and the Flex tier lists the same reduced rates for lower-priority processing with higher latency.
Related reading
GPT-6 Astra pricing, Claude API pricing, GPT-6 Astra vs Claude Fable 5.1, cheapest LLM APIs, ChatGPT Pro pricing: what do the $100, $200 and $500 tiers include?