Cheaper Flagships, Premium Speed Tiers and Spend Caps: What Three Pricing Moves Mean for Developers' AI Budgets

Background

A model’s cost to developers is often summarized as one price per million input tokens and one per million output tokens. Three developments from September and October 2026 complicate that summary: a vendor offering near-flagship results at a fraction of the flagship’s token price, vendors selling the same model at higher prices for faster output, and an employer reportedly capping what its developers may spend on tokens.

OpenAI released GPT-6 Astra on 3 September and GPT-6 Sol on 22 September, alongside a cheaper Luna model, The Next Web reports. OpenAI’s pricing page lists standard rates per million tokens of $10.00 input and $50.00 output for gpt-6-astra, and $0.10 input and $0.50 output for gpt-6-luna.

On the demand side, The Register reported on 1 September 2026 that it had seen information indicating Red Hat is limiting its developers to a maximum of $300 per calendar month in token use, with sharing of token allowances prohibited. The Register said it had asked Red Hat for comment and had received no response, and that it had no data on what Red Hat spent on tokens before the cap. The Herald’s earlier report on that item is here.

Current situation

A lower-priced model positioned against the flagship

At its DevDay event on 29 September, OpenAI launched GPT-6.1 Sol, which it says delivers nearly the same level of intelligence as GPT-6 Astra for agentic coding, computer use and professional work at one-fifth of the standard input and output token prices, according to TechCrunch. The Next Web reports API prices of $2 per million input tokens and $10 per million output tokens, with cached input at $0.10 per million tokens; OpenAI’s pricing page lists the same figures. The Herald covered the launch earlier.

The performance comparisons come from the vendor. The Next Web reports that on Terminal-Bench Science 0.1, GPT-6.1 Sol costs $5.47 per task on average at maximum effort, according to OpenAI, against $23.21 for Opus 5.5 and $23.80 for Astra, while Astra still has the highest score of the models tested, at 68.1%, and OpenAI said Astra should be used for the hardest scientific research tasks. The same report states that OpenAI ran its own models’ tests in its research environment or through its API and took competitors’ results from public reports. TechCrunch attributes the claim of near-Astra performance to OpenAI.

Gizmodo writes that OpenAI had announced it would not release GPT-6.1 Astra, that “Sol is no longer the high-powered model” it was under earlier naming, and that readers may be forgiven for losing track of OpenAI’s models. TechCrunch, citing a Wall Street Journal report, says OpenAI scrapped the Astra release over safety concerns raised by researchers during internal testing.

Faster output at a higher price

OpenAI’s Ultrafast mode documentation calls it the fastest service tier in the OpenAI API and says to use it when speed justifies the higher cost. The pricing page lists, per million tokens, the following for gpt-6.1-sol: $2.00 input and $10.00 output at standard rates, $4.00 and $20.00 in Fast mode, $12.00 and $60.00 in Ultrafast, and $1.00 and $5.00 in Batch. For gpt-6-astra it lists $10.00 and $50.00 standard, $20.00 and $100.00 Fast, and $60.00 and $300.00 Ultrafast. The page also states that priority processing was renamed Fast mode on July 30, 2026. By arithmetic on those listed figures, Fast mode is twice the standard rate and Ultrafast six times it for both models. The Herald reported the Ultrafast listing separately.

On the speed side, The Next Web reports that OpenAI says Ultrafast generates tokens up to eight times faster in Codex. The Ultrafast documentation does not state an API speedup figure. It does say Ultrafast has separate rate limits from Standard and Fast modes, and lists default Ultrafast limits for GPT-6.1 Sol of 1,000,000 tokens per minute on the Build tier, 4,000,000 on Launch and 40,000,000 on Grow.

Anthropic sells a similar trade. Its fast mode documentation says fast mode “delivers up to 2.5x higher output tokens per second at premium pricing” and runs the same model with a faster inference configuration. It is not available with the Batch API or with a Priority Tier commitment. Anthropic’s pricing page lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens at base rates, and $8 and $40 in fast mode, which is twice the base rate for up to 2.5 times the output speed. The Herald covered the move to general availability here.

Spending ceilings

The Register cites DevOps.com’s account of Gartner research: “Nearly one-quarter of technology leaders are already spending between $200 and $500 per developer each month on AI coding tokens. About 6 percent report spending more than $2,000 per developer monthly.” The Register adds that some developers on Reddit report spending from $1,000 a month to several thousand a month, and that reports across the industry vary widely. Gartner had been unable to respond immediately to The Register’s request for comment.

Implications

The following points are analysis drawn from the sources above, not reported facts.

One model, several prices. The listed figures show that a single OpenAI model now carries several per-token rates depending on service tier. For gpt-6.1-sol, the input price ranges from $1.00 (Batch) to $12.00 (Ultrafast), a twelvefold spread. Across the lineup, Astra’s Ultrafast input rate of $60.00 is 600 times Luna’s standard $0.10. A per-token comparison between vendors therefore depends on which tier is compared, and the Anthropic and OpenAI speed offers are not like-for-like, since their speed claims are stated in different terms (up to 2.5x output tokens per second versus up to eight times faster in Codex).

Task cost versus token cost. The vendor-supplied Terminal-Bench Science figures are expressed per task rather than per token, and the same report says Astra still scores highest on that test. Taken together, these suggest a developer comparing models would need cost per completed task and score on their own workload, because The Next Web states the competitor figures in OpenAI’s comparison were taken from public reports rather than re-run by OpenAI. No cited source reports an independent replication of OpenAI’s numbers.

Unit price and total spend are separate questions. The Register’s report describes a budget cap set per developer per month, which operates regardless of the per-token price of whichever model is used. The cited sources do not show whether lower per-token prices, such as Sol’s, reduce total spend or increase usage, and The Register says it has no data on Red Hat’s spending before the cap. The Gartner figures it cites describe a wide range, from $200 to $500 for about a quarter of technology leaders to more than $2,000 for about 6%.

Speed becomes an optional line item. Because Fast and Ultrafast modes are opt-in and priced at multiples of the base rate, with separate rate limits, a latency-sensitive agent loop and a background job running the same model can have very different cost profiles. Anthropic’s exclusion of fast mode from the Batch API, and OpenAI’s Batch pricing at half the standard rate, place the cheapest and fastest options at opposite ends of the same price list.

What is not known

• Independent measurements of how much faster Ultrafast or Anthropic’s fast mode runs in typical API workloads are not in the cited sources.
• Red Hat has not confirmed the cap, and The Register’s report relies on an unnamed insider.
• It is not established from these sources whether the GPT-6.1 Sol price point reflects lower serving costs or a pricing decision.