GPT-6 Pricing Rebuilt: Same List Prices as Claude, a 2.5x Gap on the Bill

GPT-6 Luna costs $0.10 per million input tokens and $0.50 per million output tokens. Claude Haiku 5.5 costs exactly the same. GPT-6 pricing now mirrors Anthropic's list prices at every tier, so the headline rate card tells you almost nothing about your bill. On one long-context agent workload we priced, two models with identical $10 and $50 list prices produced bills 2.5x apart.

The short answer

GPT-6 pricing runs from $0.10 and $0.50 per million tokens for Luna, through $2 and $10 for Sol, to $10 and $50 for Astra. Those match Claude Haiku 5.5, Sonnet 5.5 and Fable 5.1 exactly. Which vendor is cheaper depends on your cache hit rate, your prompt length and how many tokens the model spends thinking.

GPT-6 pricing, rebuilt from OpenAI's own rate card

Every figure in this section comes from OpenAI's API pricing page, read on 24 October 2026. The page carries no "last updated" date, so treat each number as a snapshot. All prices are per million tokens on the Standard tier.

The lineup arrived in two steps. GPT-6 Astra launched in early September 2026, and Artificial Analysis published its benchmark of the model on 9 September. OpenAI then announced GPT-6 Sol and GPT-6 Luna on 22 September 2026. The announcement promised "50% lower API prices for Sol and Luna compared with GPT-5.6 promotional pricing." A week later, launch coverage reports, GPT-6.1 Sol arrived with a cheaper cache.

Astra is the frontier tier, at 2.5x its predecessor

GPT-6 Astra lists at $10 input, $50 output and $1.00 for cached input. Artificial Analysis noted that this is 2.5x the $4 and $20 that GPT-5.6 Sol charged. OpenAI did not cut prices at the top. It raised them, and the published logic is that Astra spends fewer tokens per task.

Sol and 6.1 Sol share a price but not a cache rate

GPT-6 Sol and GPT-6.1 Sol both list at $2 input and $10 output. The difference is the cache. GPT-6.1 Sol charges $0.10 per million cached input tokens, half the $0.20 GPT-6 Sol charges. If your prompts repeat a long system prompt or document, 6.1 Sol is strictly cheaper at the same list price.

Luna is the cheap tier, and the GPT-5.6 rows are still listed

GPT-6 Luna lists at $0.10 input, $0.50 output and $0.01 cached. Despite talk of OpenAI replacing the GPT-5.6 lineup, GPT-5.6 Sol, Terra and Luna all remain on the card. GPT-5.6 Luna sits at $0.20 and $1.20, so on that row Luna's input price halved and its output price fell by about 58%, by our arithmetic.

GPT-5.6 Sol remains at $4 and $20. The page states its promotional pricing is available "at least through November 21, 2026." If you are still running GPT-5.6 models, you are paying double for a model OpenAI no longer leads with. That gap belongs in your migration plan, and the deadlines for older models are tracked in our calendar of API price changes and sunsets.

Three vendors, three tiers, nearly identical list prices

Side by side, the three rate cards show a pattern no vendor mentions in its launch post. OpenAI and Anthropic have converged on the same list prices at every tier. Google sits below both, with scheduled increases already printed on its own page.

Sources are the three vendors' own pricing pages. Anthropic's cache writes cost 1.25x base input for a 5-minute cache, which this table omits. Claude Opus 5.5 at $4 and $20 sits between Anthropic's frontier and mid tiers and has no exact OpenAI counterpart.

Price parity this exact is rarely coincidence. It is what a market looks like when two sellers watch each other's rate cards and decide that the headline number is a cost of entry rather than a weapon. The list price now tells you which tier you are in, not which vendor is cheaper.

The same convergence shows up one layer down. Our earlier piece on the same open model priced by four inference providers found near-identical headline rates and real differences in caching and speed. The frontier labs have arrived at the same place.

Where GPT-6 pricing diverges: caching and the 272K line

If list prices match, the bill is decided by two lines most buyers skim. The first is what a repeated token costs. The second is what happens when a prompt gets long.

Cache reads are where the vendors part ways

A cache read is an input token the provider has already processed and stored. Typically it is a system prompt, a tool schema or a document reused across requests. Every vendor discounts these, by different amounts.

OpenAI prices GPT-6 Astra's cached input at 10% of its input rate, $1.00 per million. Anthropic's pricing page states that cache hits on Claude Fable 5.1 are priced at 0.025x the base input price, which is $0.25. That is a 4x difference on the token type that dominates any agent loop. At the mid tier the gap closes: GPT-6.1 Sol and Claude Sonnet 5.5 both charge $0.10.

Anthropic charges for writing to the cache, at 1.25x base input for a 5-minute cache. The OpenAI pricing page we read shows no separate cache write column for the GPT-6 models. Our workloads below leave write costs out, which flatters Anthropic slightly on short sessions and barely matters on long ones.

The long-context thresholds are not the same line

OpenAI prices every GPT-6 model in two columns. Requests at or under 272,000 input tokens pay the short-context rate. Requests above it pay double for input and cached input, and 1.5x for output. Astra moves from $10 and $50 to $20 and $75.

Anthropic's page says the full 1 million token window comes "at standard pricing" on Claude 4.6 and later models, except Haiku 5.5. Haiku 5.5 is the exception that cuts the other way: its rates rise 5x for prompts over 100,000 tokens. Google lists no prompt-length threshold for Gemini 4 Argon or 3.8 Flash at all.

So the thresholds sit at 100,000 tokens on Anthropic's small model, at 272,000 on every OpenAI model, and nowhere on Google's two newest. A team that compares vendors at a 10,000-token prompt and then deploys at 300,000 is comparing the wrong prices. Whether you need those long prompts at all is a separate question, worked through in our cost comparison of retrieval against long context.

Workload 1: high-volume extraction on the small tier

The first workload is the job small models exist for. One million requests a month, each classifying or extracting fields from a support ticket. Each request sends 2,000 input tokens, of which 1,500 are a cached instruction block and 500 are the ticket. Each returns 200 output tokens.

That totals 1.5 billion cached tokens, 500 million fresh input tokens and 200 million output tokens a month. All workloads use Standard tier pricing with no batch discount, and none of the numbers are measured usage. They are arithmetic on published rates.

GPT-6 Luna costs $15 for the cache reads, $50 for fresh input and $100 for output, a total of $165 a month. Claude Haiku 5.5 comes to exactly $165 as well, because every rate matches. Gemini 3.1 Flash-Lite comes to $462.50, about 2.8x more, driven by its $1.50 output rate.

Output is 10% of the tokens and 61% of the Luna bill. If you are trying to cut a small-tier bill, shorter outputs and stricter schemas will do more than switching vendor. Your routing rules between small and large models matter more than the small model's brand.

The case where Luna and Haiku stop being equal

Change one variable. If each request carried a 150,000-token contract instead of a short ticket, Haiku 5.5 crosses its 100,000-token threshold and its input rate rises to $0.50. Luna stays at $0.10, because 150,000 tokens sits well under OpenAI's 272,000 line. On long documents at the small tier, Luna is 5x cheaper on input at the same list price.

Workload 2: retrieval answers on the mid tier

The second workload is a retrieval-augmented assistant answering staff questions over internal documents. It runs 100,000 requests a month. Each sends 20,000 input tokens: 5,000 of cached instructions and 15,000 of freshly retrieved passages. Each returns an 800-token answer.

All figures are Zan Digital arithmetic on the vendors' pricing pages as read on 24 October 2026. Token volumes are assumptions, not measurements. Cache write charges, explicit cache storage fees and data residency uplifts are excluded.

On this workload GPT-6.1 Sol and Claude Sonnet 5.5 tie at $3,850 to the dollar. Fresh input is $3,000 of that, because retrieved passages are new on every request and cannot be cached. Gemini 3.8 Flash comes to $1,462.50 today and $2,925 after its listed increase on 1 January 2027.

Flash is Google's mid-priced model, not necessarily a quality match for Sol or Sonnet, and the scheduled doubling means any budget built on today's rate is wrong in ten weeks. The detail of that change is covered in our breakdown of the Gemini Flash price increase. If you were still on GPT-6 Sol rather than 6.1, the same workload costs $3,900, because of the higher cache rate.

Workload 3: a long-context agent on the frontier tier

The third workload is where GPT-6 pricing and Claude pricing separate. A coding or research agent makes 20,000 calls a month. Each call carries 300,000 input tokens, mostly a repository or document set it has already seen: 270,000 are cache hits and 30,000 are new. Each call returns 4,000 output tokens.

At 300,000 input tokens every Astra request crosses the 272,000 threshold. So it pays $2.00 for cache reads, $20 for fresh input and $75 for output. The month comes to $28,800. Claude Fable 5.1 bills the same tokens at $0.25, $10 and $50, for $11,350.

$9,450 of the difference is the cache read rate. $8,000 is the long-context surcharge on fresh input and output. Neither appears in the headline $10 and $50 that both vendors advertise.

The rest of the frontier tier reorders things again. Claude Opus 5.5 at $4 and $20, with $0.20 cache hits, bills $5,080. Gemini 4 Argon bills $2,540 today and exactly $5,080 after 1 December, the same as Opus 5.5. GPT-6.1 Sol, a mid-tier model, bills $4,680 on this workload even at its long-context rates, and Claude Sonnet 5.5 bills $2,540.

The threshold is the lever. Hold everything else and cut each prompt to 270,000 tokens, 240,000 cached and 30,000 fresh. Astra then stays on short-context rates and the month costs $14,800. Fable 5.1 on the same trimmed prompt costs $11,200. The gap shrinks from 2.5x to about 1.3x. On OpenAI, context discipline is worth more than any discount you could negotiate.

Where this comparison is weakest

Every number above is correct arithmetic on published rates. That is a narrower claim than "this is what you will pay," and three gaps sit between them.

Output volume is not fixed in real life

Our workloads hold output at 4,000 tokens per call for every model. Real reasoning models do not behave that way. Artificial Analysis measured GPT-6 Astra using 27,000 output tokens per task at max effort, against 78,000 for Claude Fable 5.1. On its Intelligence Index, Astra matched Fable 5.1 at about 40% of the cost per task.

That is the strongest case against our headline. On a reasoning-heavy benchmark with short prompts, Astra is the cheaper frontier model, and by a wide margin. Our 2.5x gap only holds where prompts are long, heavily cached and output volumes are similar. If your workload looks like a benchmark rather than an agent re-reading a repository, the ranking may flip.

Per-token parity is not per-page parity

Identical prices per token only mean identical prices per page if the tokenizers count text the same way. They do not, and nobody publishes a cross-vendor conversion. Anthropic's own pricing page warns that Claude 4.7 and later models use a newer tokenizer that "produces approximately 30% more tokens for the same text" than its earlier one.

That comparison is against Anthropic's previous tokenizer, not against OpenAI's, so it does not tell you which vendor is cheaper per page. It does tell you that a per-token tie can hide a real difference. Count tokens on a sample of your own documents with each vendor's counter before you trust any per-token comparison, including ours.

The tier mapping is a judgement

We matched models by their position on each card. Gemini 4 Argon is Google's newest model and, from 1 December, its most expensive listed text model, but that does not make it a capability match for Astra or Fable. Gemini 3.8 Flash may not match Sol on your tasks either. A cheaper model that fails 10% more often is not cheaper, and our sums say nothing about quality.

We also left out Claude Opus 5.5 as a direct pair, because OpenAI has no $4 and $20 model in the GPT-6 family. For many teams Opus 5.5 is the real alternative to Astra, and at $5,080 on workload 3 it is the comparison worth testing.

The multipliers that sit on top of every rate card

The Standard tier is the starting point, not the bill. All three vendors publish multipliers for speed, delay and location, and they are surprisingly aligned.

OpenAI's Fast multiplier is our arithmetic from the table values; the page states the prices, not the ratio. OpenAI notes that Priority processing was renamed Fast mode on 30 July 2026.

Batch is the one lever that pays everywhere. Any workload that can wait hours, such as overnight extraction or evaluation runs, halves its cost on all three vendors. Workload 1 drops to about $82 a month on Luna or Haiku.

Speed is where the expensive surprise lives. Astra's Ultrafast tier costs 6x Standard. We wrote about what OpenAI's Ultrafast tier actually buys; the short version is that a latency premium belongs on the user-facing slice of traffic, never on the whole account. And if the token price keeps falling while the bill keeps rising, the reason is usually volume, which is the subject of our analysis of cheap tokens and rising AI bills.

Frequently asked questions

Where to start: re-price your top workload this week

Take the single workload that costs you most. For a typical request, write down four numbers: cached input tokens, fresh input tokens, output tokens and total prompt length. Then price that request on all three rate cards with the thresholds applied, not just the headline rates.

If the prompt length lands between 100,000 and 272,000 tokens, or above 272,000, you have found the line where vendor choice actually matters. Run a two-week test on the cheapest qualifying model before you renegotiate anything. The list prices have converged, and we expect them to stay converged; the savings are now in how you shape the request.

Related reading