OpenAI Core Product Head Tibo: gpt-6.1-sol Becomes the Most Demanded Model in Company History
OpenAI Core Product Head Tibo Sottiaux stated that gpt-6.1-sol has become the most demanded model in the company's recorded history, with pressure on both subscription and API sides.
The model was launched on September 29 during devday, positioned as an upgraded version of gpt-6 sol. The official claim is that it is close to gpt-6 astra in terms of agentic coding, computer use, and professional work, with standard input-output pricing about one-fifth of astra. API pricing is $2 per million tokens for input, $0.10 for cached input, and $10 for output; the cached read price has been halved from gpt-6 sol's $0.20, and costs decrease significantly when the long context agent repeatedly returns the same prompt.
The product entry is constrained to paid workflows: plus, pro, business, enterprise, and edu can use it in chatgpt work and codex, while the regular chat window is not yet open, and free and go are not included. The API model ID is gpt-6.1-sol, with a context window of 1,050,000 tokens and a maximum output of 128,000 tokens; codex cli 0.159.1 has set it as the default model, while github copilot is gradually rolling it out to pro+, max, business, and enterprise.
Demand is being driven from two pipelines simultaneously. Subscription users are pushing their quotas into long tasks in codex and work, while developers are switching their default calls from gpt-6 sol to gpt-6.1-sol, as the model reports 75.2% on deepSWE v1.1 under high effort, surpassing gpt-6 sol's 68.8% at max effort, and matching astra, with single-task costs about one-fifth of astra. The plus tier estimates local messages for gpt-6.1 sol at about 15 to 160 per five hours, close to gpt-6 sol's 15 to 150, but with longer single tasks and more tool calls, instantaneous computing power consumption is amplified.
Tibo explicitly stated that chatgpt and codex were previously under heavy load, and the company has brought more capacity online, claiming that speeds will significantly rebound in the coming hours, approaching nearly double yesterday's service levels. The official also previously announced that gpt-6.1 sol ultrafast will be available in the coming days, with codex side token generation up to about 8 times the standard tier; current congestion occurs just as the standard tier has become the default and ultrafast has not yet fully ramped up.
In terms of market mechanics, the buyers are paid subscription users and API developers billed by token, while the sellers are OpenAI's inference capacity. This is a typical event-driven scenario: after devday, prices remain unchanged, capabilities increase, and codex defaults switch, causing the demand curve to shift outward instantly, while supply cannot keep up in the short term. Funds are shifting from the $10/$50 tier of astra and competing coding agents to the $2/$10 sol tier; OpenAI gains higher call volumes, but in the short term, delays and quota experiences are under pressure, while astra faces the challenge of being "more expensive yet no longer the default choice."
On a supplementary level, this model has appeared in distribution layers such as openrouter, vercel ai gateway, and azure, with multiple entries overlapping the same cluster. The official has not disclosed specific QPS or queue seconds, only stating "expanded capacity, approaching nearly double yesterday's speed within hours" as an operational metric.
Source: Public Information
ABAB AI Insight
Tibo Sottiaux previously made codex one of OpenAI's fastest-growing revenue streams, placed above both chatgpt and codex by mid-2026, reporting directly to Greg Brockman. His approach is not to launch a flagship first and then cut down, but to make sol the default workhorse: during the gpt-5.6 sol period, he explained the relationship between cached read/write, context over 272k, and subscription billing; this time, 6.1 sol follows the same logic—prices remain unchanged, caching is halved, and capabilities align with astra. Internally, the company is synchronously shrinking independent products like sora and channeling traffic into a "super application"; once the codex default model switches, the entire company's workflow will hit the same inference bottleneck.
The capital path is clear: sell the intelligence of astra's $10/$50 at $2/$10 and $0.10 cached reads to high-frequency agents. The motivation is to occupy the developer default position, rather than sell a more expensive flagship ticket. Resource mobilization occurs in three areas: setting gpt-6.1-sol as the default for codex cli, locking subscription quotas in work/codex rather than regular chat, and treating ultrafast and the $500 pro 500 as premium pressure relief valves. Money shifts from "occasional flagship calls" to "continuous mid-tier model usage," with OpenAI trading usage for gross profit and expansion for retention.
Analogies can be drawn to Google making gemini flash the default, Anthropic using sonnet to block opus, and AWS pushing graviton as the cost-effective default core. OpenAI is currently in a control phase: not expanding another model line, but consolidating entry points, defaults, and quotas onto the same sol framework. Historical parallels include gpt-4o being squeezed out of daily use by free traffic and gpt-3.5 turbo washing the completion market into a bulk API. The position of 6.1 sol is closer to being the "production default core," while astra slides towards being a backup option that gets stuck before switching.
Structural judgment indicates a transfer of pricing power. The mechanism is: when mid-tier models are close to flagship performance on real codebase tasks, and caching discounts collapse multi-round agent costs, buyers no longer pay five times the price for marginal intelligence, and sellers can only reprice based on capacity and speed layers. Congestion itself is evidence that the transfer is complete—the demand has collectively shifted to the sol tier, and flagship pricing power has been dismantled by usage frequency.
ABAB News · Cognitive Laws
Default items determine cash flow more than flagships.
Price cuts without increased capacity equate to selling congestion to users.
Once mid-tier models align with flagships, the expensive ones become backup parts.