The Bulk Discount Is Gone: OpenAI's Pro 500, the Pro 200 Cut and GPT-6.1 Sol on Real Work

OpenAI’s DevDay shipped a lot on September 29. Three announcements change what heavy users pay: a $500 Pro plan, a smaller allowance on the $200 plan, and GPT-6.1 Sol. The night before, Thibault Sottiaux of OpenAI’s Codex team explained the plan change in advance.

In effect, if you do the math, it will net out at half the dollar in API spend compared to the old Pro $200 plan.

— Thibault Sottiaux, OpenAI

What Changed

• Pro 500. OpenAI’s help page calls it “a new $500/month plan that includes Astra Ultrafast”. It is the only plan with Ultrafast.
• Pro 200 reopens with less. New subscribers get “a lower usage allowance than previously offered with Pro 200 to reflect our increasingly efficient models”. Existing subscribers keep the old allowance through October 29.
• GPT-6.1 Sol. OpenAI says it “nearly matches GPT-6 Astra’s intelligence” at one-fifth of Astra’s price: $2 in, $10 out, and cached input down to $0.10.
• Ultrafast. A service tier, not a model: up to 8x faster in Codex at six times the API price. It is live for Astra, and Sottiaux says it is coming soon for 6.1 Sol.

The Bulk Discount Is Gone

OpenAI’s help page gives no multiples. The email to existing Pro 200 subscribers, which I received, does: “Starting October 30, 2026, your included usage in ChatGPT Work and Codex will decrease from 20 times to 10 times the ChatGPT Plus allowance.” Pro 500 is “25 times the usage of Plus”. Existing subscribers also get a one-time grant of 62,500 usage credits, “a $2,500 USD value”, which expires on December 31.

Every paid tier now buys the same usage per dollar. The old Pro 200 bought twice as much. Paying more used to get you a better rate. Now it only gets you more of the same rate, plus speed.

Sottiaux says this is the goal. He wants prices low enough that “it makes sense for most to buy usage as needed without there being a significant gap between what you get in a subscription and what you get in the API for a dollar spent”. A subscription is becoming prepaid API credit.

What $200 Used to Buy

I pay $200 a month for Codex and $200 each for three Claude Max seats. Priced at API list rates from my own logs, the last 30 days came to:

• Claude, three seats: about $11,300 on $600, roughly 19x. That excludes this week’s benchmark runs.
• Codex: at least $3,200 of GPT-6 Astra on $200, at least 16x. About 1,400 older GPT-5.6 Sol turns are excluded because I can’t price them reliably.

SemiAnalysis estimated in June that a fully used Pro 200 was worth about $14,000 a month at API prices. OpenAI has known this for a long time.

insane thing: we are currently losing money on openai pro subscriptions!

— Sam Altman, January 2025

Halving the allowance does not end the subsidy. A heavy user still gets many times the price. It ends the part of the subsidy that grew with the plan size.

Speed Is the New Premium

Pro 500 is not a volume plan. At 25x it buys the same usage per dollar as Plus. The email is explicit that what it adds is speed: Ultrafast, “with up to 8 times faster token generation”. That speed runs on scarce hardware. In August, Cerebras said it was “powering Ultrafast mode” for GPT-5.6 Sol. Neither company has said who serves the new tier.

Capacity was already tight. According to Fortune, OpenAI paused Pro 200 sign-ups on September 10 when Astra demand strained it. It reopened them 19 days later at half the allowance.

GPT-6.1 Sol on Real Work

I ran 6.1 Sol through the same bench as this week’s effort-level post: six real merged changes from a production monorepo, graded by the tests the original fixes shipped with. Medium effort, 20 runs, list prices. I reran GPT-6 Sol on the same day for comparison.

• It is smarter. It fixed more new tests than any model at medium, and it got three clean runs on a change where the others got one or none.
• It is the cheapest. $6.73 for 20 runs, 57% of Sonnet 5.5 and 15% of Opus 5.5.
• It is still careless. On the change where Sol models rewrite an existing message and drop the instruction it carries, 6.1 Sol did it in all four runs. GPT-6 Sol did it in three.

Is This the End of $200 Heavy Use?

For OpenAI, yes. The top plan no longer buys more per dollar, and the new top tier sells speed. OpenAI’s API price cuts continue, so the tokens you get from a fixed allowance may still hold up. But the plan no longer pays you for committing more money.

Anthropic went the other way last week. The Opus 5.5 launch raised five-hour limits by 20% and said Opus “goes 25% further” within them. It has said nothing about following OpenAI.

Both companies are heading toward listings. Anthropic’s leaked prospectus shows $7.33 billion of compute spend in 2025. One Hacker News commenter read OpenAI’s timing as cleaning up the numbers “ahead of Q4 so that their financials will look better going into next year for the IPO”. That is a guess, but the incentive is real for both labs.

What This Doesn’t Show

• The multiples are relative. OpenAI states allowances as multiples of Plus, not in tokens or dollars, so “half the dollar” is Sottiaux’s conversion, not a published figure.
• List-price value is not cost. My 16x and 19x measure what the usage would cost on the API, not what it costs OpenAI or Anthropic to serve.
• One bench, 20 runs. Most gaps in the table are two or three runs.
• Ultrafast is untested here. It is not on Pro 200, and Sol does not have it yet.

What I’m Doing With It

• Keeping Codex on Pro 200 through October 29, then checking whether the halved allowance covers my use.
• GPT-6.1 Sol for well-tested work. It is the cheapest model on my bench and the best at the task itself. A test suite catches its habit.
• Claude for work without that safety net. Zero regressions in 40 runs across both models at medium.

The $200 plan used to reward the heaviest users with the best rate. OpenAI just made the rate flat. The price of a token keeps falling, but the discount for buying a lot of them has ended.