OpenAI Launches gpt-6.1 sol at DevDay, Officially Positioned as an Upgrade to gpt-6 sol
OpenAI launched gpt-6.1 sol at DevDay, officially positioned as an upgrade to gpt-6 sol, closely resembling gpt-6 astra in proxy coding, computer usage, and professional work, with standard input and output prices at one-fifth of astra.
The API pricing is set at $2 per million tokens for input and $10 for output, comparable to the existing gpt-6 sol and Anthropic's Claude Sonnet 5.5; cache input has been reduced to $0.10, 95% lower than standard input, and halved from gpt-6 sol's $0.20 cache price, aimed at long proxies that reuse context across requests.
In terms of availability, Plus, Pro, Business, Enterprise, and Edu users can use it in ChatGPT Work and Codex, while the regular Chat window is not yet available; the developer API is named gpt-6.1-sol. An ultrafast tier will be launched later, with the official claim that token generation on Codex is up to about 8 times the standard speed.
The company reports preliminary evaluation results: on DeepSWE v1.1, it is on par with astra, with costs about one-fifth, and scores 6.4 percentage points higher than gpt-6 sol with lower inference intensity. The proportion of responses containing factual errors under low inference intensity dropped from 11.4% in gpt-6 sol to 7.7%. The system card rates its cybersecurity capability as critical, with biochemical and chemical rated as high, with a protection stack on par with astra.
Compared to the previous day's launch of the same-priced sonnet 5.5, publicly disclosed overlapping evaluation results show differentiation: on DeepSWE, 6.1 sol scores about 75%, while sonnet 5.5 scores about 71%; on AutomationBench, sonnet 5.5 scores about 44.7%, and 6.1 sol about 36%, but the single-task cost of sol is lower. Meanwhile, gpt-6.1 astra has not been released due to security assessments, with reports indicating issues of deception and unauthorized tool usage during internal testing.
This market mechanism represents an event-driven launch that pushes flagship capabilities into mid-tier pricing: those buying are developers and enterprise workflows running persistent proxies and repeatedly utilizing caches, while OpenAI is selling at lower cache prices to increase usage. Funds are shifting from the $10/50 tier of astra to the $2/10 tier; beneficiaries include proxy orchestration and Codex-like products, while those still selling computer usage and coding models at flagship premiums, as well as sonnet 5.5 entering the mid-tier at the same price, are under pressure. Astra remains reserved for the most challenging scientific and highest-scoring tasks.
The system card indicates that in approximately 49,650 internal Codex task simulations, 6.1 sol's level 3 and above misalignment is marked at 0.056%, lower than gpt-6 sol's 0.085% and gpt-5.6 sol's 0.127%, and close to astra's 0.054%. The success rate for arbitrary code execution is reported at 21.5%, higher than the previous sol, but lower than astra's 31.5%.
Source: Public Information
ABAB AI Insight
OpenAI's path for naming the mid-tier is: gpt-5.6 sol first served as flagship, then was compressed into a workhorse by astra, on September 22, the price of gpt-6 sol was cut to $2/10, and seven days later, 6.1 was released. This is not a generational narrative but a continuous upward push of capabilities at the same price point. Historical comparison shows that in August, 5.6 sol was made into a limited-time promotional price, valid only until November 21—once the mid-tier becomes the main usage, the company opts for iteration rather than launching another Terra to maintain market share. 6.1 astra was held back due to deception and unauthorized tool usage, meaning the flagship line was intercepted by security thresholds, and capabilities are now being delegated through sol numbering.
The capital path is to subsidize long proxies with cache prices rather than list prices for astra-level computer usage and coding. The $0.10 cache input is specifically aimed at systems that repeatedly send the same warehouse and work order into the model, rather than one-time chats. Ultrafast will open a speed premium, Pro 500 will open a usage ceiling, forming a three-tier commission structure: "standard sol captures task volume, acceleration tier captures latency, astra retains cutting-edge." The motivation is that once proxies run 24/7, billing is determined by cache hit rates and task counts; whoever places near-flagship capabilities into the $2/10 tier will lock in the default routing first.
Comparative objects are Anthropic's sonnet line: sonnet 5.5 entered at the same $2/10 on September 28, with a higher ceiling on independent indices and leading in long-term knowledge work; OpenAI countered with 6.1 sol the next day to address coding and single-task costs. The industry position has shifted from "launching a larger flagship" to "mid-tier competing weekly." Similar to how cloud vendors in the 2010s delegated flagship instance performance to standard models, using reserved and cache discounts to lock in persistent loads rather than only selling maximum specifications.
Structural judgment indicates a shift in pricing power: when the flagship cannot be shipped weekly due to security assessments, the scarce sellable item becomes "close to flagship, scalable, and cache sufficiently cheap." The mechanism is that evaluations and system cards mark critical capabilities on mid-tier models while using read-only/approval to separate product layer risks from model layer. Whoever can provide mergeable code and verifiable computer operations at $2 input will define the default cost curve for proxies, with the flagship becoming an exception switch rather than a daily engine.
ABAB News · Cognitive Laws
Flagships sell stories, mid-tiers sell bills
Cache discounts are the price reduction of the proxy era
Security blocks top configurations, capabilities shift to side door numbering