Claire Donovan
OpenAI released a model this week that it barely bragged about. It is not the fastest, it does not top independent leaderboards, and it is not even in the regular ChatGPT app. Yet GPT-6.1 Sol may do more to change who gets to build with artificial intelligence (AI) than any flagship launch this year, because it makes near-frontier intelligence cheap enough to run all day. (See also: The AI That Broke Out)
Announced September 29, 2026 at DevDay in San Francisco, GPT-6.1 Sol is positioned as near-GPT-6 Astra intelligence at one-fifth of Astra's standard input and output prices ((The Hindu, 2026)). The real story is not a benchmark crown. It is a pricing bet on who gets to use professional-grade AI all day.
What Is GPT-6.1 Sol?
GPT-6.1 Sol is a text-and-image model with a 1,050,000-token context window, a 128,000-token maximum output, and an April 30, 2026 knowledge cutoff ((OpenAI, 2026)). It is available through the application programming interface (API) as gpt-6.1-sol and inside ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, but not in standard Chat mode ((IT Brief, 2026)). BenchGecko's normalized multi-benchmark score for GPT-6.1 Sol is 93.4, ranking it sixth out of 316 scored models, in the top 2% ((BenchGecko, 2026)). That sounds elite. The catch is that near-frontier does not mean frontier.
It also supports text output only, not audio or video generation, and lacks fine-tuning, realtime, embeddings, or image generation ((OpenAI, 2026)). So Sol is not a drop-in replacement for Astra across multimodal and specialized workloads.
How Does GPT-6.1 Sol Compare to GPT-5 and GPT-6?
If you are waiting for a GPT-5 comparison, this is not that launch. Sol slots into the GPT-6 family, above the older GPT-6 Sol and just below the flagship GPT-6 Astra. Independent testing from Artificial Analysis places GPT-6.1 Sol at 52 on its Intelligence Index, one point below GPT-6 Astra (53) and four points above GPT-6 Sol (48) ((DeepLearning.AI, 2026)). Anthropic's Claude Sonnet 5.5 and Claude Opus 5.5 led the index at 56 and 58 ((DeepLearning.AI, 2026)).
In OpenAI's self-reported DeepSWE v1.1 software-engineering benchmark, Sol matched Astra at roughly one-fifth the cost per task and beat GPT-6 Sol's best score by 6.4 points ((IT Brief, 2026)). That is a meaningful jump for a cheaper model. But several benchmark claims are OpenAI-reported and deserve independent replication beyond the Artificial Analysis composite.
Why Is OpenAI Launching Near-Frontier AI at a Fifth of the Price?
OpenAI's list pricing is aggressive: $2.00 per 1M input tokens, $10.00 per 1M output tokens, and $0.10 per 1M cached input tokens, versus Astra's $10.00, $50.00, and $1.00 ((OpenAI, 2026)). Forkast.News (2026) called it the sharpest cost compression in the frontier tier to date and framed it as a step toward commoditization of frontier intelligence.
'very near-Astra-level intelligence at a fifth of the price' - Sam Altman, via The Hindu (2026)
That is not charity. The move is widely read as a bet on AI agents, whose autonomous multi-step work drives token consumption sharply higher. Zovora's own reporting on 79% of Companies Are Using AI Agents Right Now - But Only a Fraction Actually Have Them Working shows the same demand wave. Cheap reasoning becomes a volume play.
When and Where Can You Access GPT-6.1 Sol?
Sol is available today to developers via the API as gpt-6.1-sol and inside ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users ((DeepLearning.AI, 2026)). If you already pay for ChatGPT Plus or Pro, you do not need a new subscription to use Sol inside Work and Codex ((IT Brief, 2026)). That means near-flagship help arrives as part of an existing plan, not as a five-times-more-expensive add-on.
OpenAI also introduced a Pro 500 tier at $500 per month for heavy users, while casual users can stay on cheaper plans ((Economic Times, 2026)). Standard ChatGPT Chat mode still does not include Sol, so do not expect it in the regular chat window yet ((IT Brief, 2026)).
How Much Does GPT-6.1 Sol Cost and Is It Worth It?
Here is the number worth telling someone: reading 10 million cached input tokens costs $1.00 on Sol versus $10.00 on Astra ((OpenAI, 2026)). For a coding agent re-reading a large codebase or a document tool looping through contracts, repeated multi-step tasks drop from dollars to cents. Artificial Analysis measured the index completion cost at $0.72 per task for Sol versus $3.26 for Astra, and found no cheaper model at Sol's performance level at any tested reasoning setting ((DeepLearning.AI, 2026)).
On Terminal-Bench Science 0.1, Sol averaged $5.47 per task at maximum reasoning effort versus $23.21 for Claude Opus 5.5 and $23.80 for Astra, though Astra still posted the highest score at 68.1% ((IT Brief, 2026)). For a bootstrapped founder, that is the difference between an API bill and product development, a theme Zovora explored in I Built an AI Product for $400/month Instead of $40,000 - Here's What Broke.
Who Should Use GPT-6.1 Sol?
Sol is built for entry-level builders. A 22-year-old computer science student debugging a side project in Codex can now see agentic coding tasks that once cost over $20 per run drop to roughly $5.47 per task at maximum effort ((IT Brief, 2026)). That makes all-night iteration affordable on a student budget. A junior developer already paying for ChatGPT Plus gets near-flagship Sol inside Work and Codex without upgrading ((IT Brief, 2026)).
The deeper shift is skill-based: routine implementation moves to agents, and the human job becomes specifying tasks, reviewing output, and debugging failures. That is the same pressure point mapped in Zovora's story Apple vs OpenAI: The Lawsuit That Just Changed Every Tech Job In 2026. (See also: Apple vs OpenAI)
Why Astra and Claude Still Matter
Near-frontier is not frontier. On Artificial Analysis' Intelligence Index, Sol at 52 still trails Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56, and even GPT-6 Astra at 53 ((DeepLearning.AI, 2026)). Speed matters too: Sol managed 57.7 tokens per second, about two-thirds of GPT-6 Sol's 87.6 ((DeepLearning.AI, 2026)). For interactive or latency-sensitive apps, that gap is real.
Sol also does not support audio, video, fine-tuning, realtime, embeddings, or image generation, so it is not a drop-in replacement for Astra across multimodal and specialized workloads ((OpenAI, 2026)). Separate from this launch, Zovora has covered how autonomous agents expose security blind spots in The AI That Broke Out. The cheap model story coexists with unresolved questions about agent autonomy.
If you are already paying for ChatGPT Plus or Pro, try Sol inside Work or Codex before upgrading. If you are building agents, run the math on cached input at $0.10 per 1M tokens. If you need the hardest reasoning, keep Astra or Claude in the loop. Cheap enough is not the same as best, but for a generation of builders, it may finally be good enough to run all day.
