ChatGPT Pro 500 Plan and GPT-6.1 Sol Pricing Explained

OpenAI's DevDay shocker: a $500 subscription and a frontier-class model at one-fifth the price

OpenAI used its 2026 DevDay on September 29 to restructure both its subscription ladder and its model lineup in a single stroke. The company introduced a $500-per-month ChatGPT Pro 500 tier — its most expensive consumer plan ever — while simultaneously launching GPT-6.1 Sol, a model it claims approaches the capability of its flagship GPT-6 Astra at one-fifth of the token price. The dual announcement came barely a week after GPT-6 Sol shipped on September 22, and just as OpenAI reopened a $200 Pro plan it had frozen on September 10 amid capacity strain from unprecedented demand for GPT-6 Astra.2228

The Pro 500 plan carries the highest usage allowance OpenAI has ever offered an individual subscriber and is the only Pro tier that includes the new "Ultrafast" processing mode, which OpenAI says generates up to 300 tokens per second in Codex — as much as eight times the standard rate.22628 But the new tier landed with a sting for existing customers: OpenAI also cut the value of the $200 plan, dropping its Codex and Work allowance from 20x to 10x the Plus baseline and halving the weekly cap on GPT-6 Pro messages from 200 to 100, effective October 30.2228

What GPT-6.1 Sol actually costs

The headline API rates are simple: $2 per million input tokens, $10 per million output tokens, and a cached-input rate of $0.10 per million — a 95 percent discount on fresh input and half of GPT-6 Sol's $0.20 cache price.1330 The input and output rates are identical to GPT-6 Sol, which means the accuracy and capability gains come at no additional token cost.225 Prompts exceeding 272,000 input tokens trigger a long-context repricing of $4 input / $15 output applied to the entire request, a structural detail that matters enormously for agent workflows, which tend to carry heavy context.2637

Against GPT-6 Astra — priced at $10 input and $50 output per million tokens — Sol's $2/$10 is exactly one-fifth the sticker.65 The model carries a 1.05 million-token context window with 128,000 maximum output tokens, and its model identifier is gpt-6.1-sol.304 Batch and Flex processing halve the rates to roughly $1/$5, while Fast mode doubles them.253

Availability is deliberately scoped: GPT-6.1 Sol is live in Codex and ChatGPT Work for Plus, Pro, Business, Enterprise, and Education users, plus the API — but not yet in the standard ChatGPT chat interface.2293

Pro plans now come in three sizes

Alongside Pro 500, OpenAI's Help Center now lists a three-tier Pro ladder: Pro 100 at $100 per month, Pro 200 at $200, and Pro 500 at $500, with monthly billing only.2425 Pro 500's allowance is described as 25 times the standard ChatGPT Plus allocation, and Ultrafast draws down that allowance at eight times the standard rate.2125 Existing Pro 200 subscribers who were active at the eligibility cutoff keep their old allowance through October 29 and receive a one-time grant of usage credits worth $2,500, expiring at year end — a compensation package that reads as an acknowledgment of how sharp the downgrade is.22

The pricing logic is blunt. SemiAnalysis estimated in June that a $200 ChatGPT Pro plan pushed to its limits with long-running coding and agent tasks would consume roughly $14,000 a month of usage at OpenAI's standard API rates, versus about $8,000 for Anthropic's Claude Max 20x at the same subscription price.22 OpenAI's answer to that economics problem is segmentation: charge the heaviest users more, cap the mid-tier, and sell speed as a premium ingredient. The move also follows industry precedent — OpenAI broke the $200 ceiling in late 2024, and Anthropic and Perplexity matched it — but $500 is a new frontier, and OpenAI is betting a segment of power users will pay $6,000 a year for capacity.2228

All Pro subscribers, regardless of tier, will receive a single "Dot," OpenAI's new always-on personal AI agent, and conversations with the Dot will not count against usage limits. OpenAI also partnered with 16 companies including Notion and Cognition on a "Sign in with ChatGPT" option that lets customers run OpenAI models inside those services without configuring API keys.2224

The benchmarks OpenAI wants you to see

GPT-6.1 Sol's launch claims rest on three evaluations. On OSWorld 2.0, a computer-use benchmark, it outperformed GPT-6 Sol by seven percentage points at maximum reasoning effort while costing less than half as much, and came within 2.1 points of GPT-6 Astra at roughly one-seventh the per-task cost.24 On Terminal-Bench Science 0.1, it more than doubled GPT-6 Sol's score at maximum reasoning effort, averaging $5.47 per task against $23.21 for Claude Opus 5.5 and $23.80 for GPT-6 Astra — though Astra remained the highest-scoring model on that evaluation at 68.1 percent.2413 Factuality improved too: on one deliberately difficult evaluation at low reasoning effort, the share of responses containing at least one factual error fell from 11.4 percent with GPT-6 Sol to 7.7 percent.24

The one caveat worth flagging: token-price ratios are not task-cost ratios. A one-fifth sticker price does not guarantee one-fifth spend per completed task, because different models reason for different lengths, retry differently, and call tools at different rates.78

Notably absent: GPT-6.1 Astra

The model OpenAI did not ship may matter as much as the one it did. A GPT-6.1 version of Astra had been expected in October, and reporting indicates it was scrapped over safety concerns after internal testing.2237 OpenAI's own evaluations still crown GPT-6 Astra, released September 3, as its most capable model.37 For enterprise buyers, a cancelled release at the frontier is a roadmap risk in a way a delayed one is not.

Anthropic's answer was already on the table

The competitive backdrop explains some of OpenAI's urgency. Anthropic shipped Claude Opus 5.5 on September 22 — its most capable Opus, with a 1M-token context window, 40 percent lower running cost than Opus 5 on typical workloads, and API prices of $4/$20 per million tokens.17 Claude Sonnet 5.5 followed on September 28, running more than 30 percent faster than Sonnet 5 at the same $2/$10 rates as GPT-6.1 Sol.1737 Independent leaderboard tracking places Claude Opus 5.5 at the top of the intelligence index at 57.6, ahead of Claude Fable 5.1 (53.4), GPT-6 Astra (52.7), Gemini 4 Argon (52.6), and GPT-6.1 Sol at 51.8 — at a measured $0.72 per index task, Sol's position on that leaderboard is the best value among the frontier names.37 Anthropic also filed an S-1 on September 28 revealing $518 billion in cloud and infrastructure commitments, roughly 80 percent non-cancelable, against $42 billion in losses and a valuation target above $2 trillion.23

Google gates its frontier, and the open-weight crowd moves in

Google announced Gemini 4 Argon on September 30 — the first Gemini 4 model, with output up to one million tokens, level with GPT-6 Astra on independent tracking — but released it only to vetted cybersecurity experts through its Fairwind Program, withholding general availability over misuse concerns.363735 Its introductory pricing is $2/$10 per million tokens against a standard $4/$20, with no GA date published.37 Google also shipped Nano Banana 2.1 on October 6 and continues a near-monthly Flash cadence.3139 Meanwhile the open-weight tier keeps pressure on everyone: Xiaomi's MiMo-V2.6-Pro (MIT-licensed, released September 21) scores 46.3 on the same index at roughly $0.13 per task — the best measured intelligence per dollar on the board — and Mistral released Mistral Large 4 on October 6, with open weights promised.373139 Google's Gemma 4 family, released April 2026, spans open-weight models from E2B to 27B.3438

The reading

Taken together, the week's news describes an industry that has stopped competing on raw capability alone and started competing on cost-per-task, speed tiers, and access control. OpenAI is segmenting demand three ways — charging heavy users $500, halving the mid-tier's compute, and selling 8x speed as an ingredient — while pushing the workhorse model's price to levels that undercut its own flagship by 80 percent.2223 Anthropic matched the price points with Opus 5.5 and Sonnet 5.5; Google kept its flagship locked behind a vetted-access program rather than price it against rivals.1736 When frontier launches get cancelled over safety tests and flagship models ship only to vetted defenders, the planning lesson is clear: rate cards now change faster than deployment cycles, and the durable skill is routing — measuring dollars per completed task on your own workload, exploiting cache discounts like Sol's halved $0.10 rate, and not assuming that any announced model will actually arrive.377