Author(s): Decoding AI by Nueravi

Originally published on Towards AI.

GPT-6.1 Sol Shipped. Most AI Frameworks Still Don’t Know It.

Key takeaways

• GPT-6.1 Sol launched on 29 September 2026. Twelve hours later, 3 of 10 widely used AI libraries recognised it: Pydantic AI, LiteLLM and the Vercel AI SDK.
• LlamaIndex 0.8.2 raises ValueError: Unknown model 'gpt-6.1-sol' the moment you read the model's metadata. It handles GPT-6 Sol correctly.
• LangChain returns an empty model profile for GPT-6.1 Sol in Python; in JavaScript it returns an empty one for GPT-6 Sol too.
• GPT-6.1 Sol drops reasoning.effort: none. GPT-6 Sol accepted it. The three libraries that know the new model already encode the change.
• The 5 libraries that misread GPT-6 Sol on 26 September still misread it on 30 September, eight days after that launch.

TL;DR

Does your AI framework support GPT-6.1 Sol yet? Twelve hours after OpenAI released it, 7 of the 10 libraries we probed did not. On 30 September 2026 at 04:56 UTC we installed the latest release of each and asked the question it asks internally before every call — how big is this model’s context, what does it cost, is it a reasoning model — for gpt-6.1-sol, with gpt-6-sol as the control.

Your API calls will mostly still go through. What breaks is everything the library does around the call: context sizing, cost tracking, token counting and which parameters it sends.

Status, 30 September 2026, 04:56 UTC: pydantic-ai-slim 2.52.0, litellm 1.103.1, @ai-sdk/openai 4.0.82, llama-index-llms-openai 0.8.2, langchain-openai 1.6.6, @langchain/openai 1.6.0, crewai 1.15.23, tiktoken 0.14.0, dspy 3.4.0, openai-agents 0.22.3. Re-check the hardest failure in one line: python -c "from llama_index.llms.openai.utils import openai_modelname_to_contextsize as f; print(f('gpt-6.1-sol'))".

What is GPT-6.1 Sol?

GPT-6.1 Sol is OpenAI’s upgrade to GPT-6 Sol, released on 29 September 2026 at $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. OpenAI’s model page lists a 1,050,000-token context window, 128,000 maximum output tokens and the API name gpt-6.1-sol. OpenAI says it nearly matches GPT-6 Astra on agentic coding at a fifth of Astra's price. Its cached-input price is half GPT-6 Sol's, which is why a cost table that treats the two as the same model gets your bill wrong.

Which AI frameworks support GPT-6.1 Sol?

Three of the ten libraries we probed recognised GPT-6.1 Sol twelve hours after launch: pydantic-ai-slim 2.52.0, LiteLLM 1.103.1 through its downloaded price list, and @ai-sdk/openai 4.0.82. Pydantic AI and the AI SDK both shipped a release in the hours after launch. Here is how each library answered for gpt-6.1-sol:

• Pydantic AI 2.52.0: known model, with a profile that rejects effort: none. Correct.
• LiteLLM 1.103.1: $2 input, $0.10 cached, $10 output, none unsupported, but only from the list it downloads at import.
• Vercel AI SDK (@ai-sdk/openai 4.0.82): reasoning model, efforts low to max. Correct, by a rule for any GPT-6 or later name.
• LlamaIndex 0.8.2: ValueError: Unknown model.
• LangChain (Python) 1.6.6: empty model profile.
• LangChain.js (@langchain/openai 1.6.0): empty model profile.
• crewAI 1.15.23: a 6,963-token context window.
• tiktoken 0.14.0: KeyError, no tokenizer mapping.
• DSPy 3.4.0: not treated as a reasoning model.
• OpenAI Agents SDK 0.22.3: reasoning settings not applied.

A model is not supported when the first request returns 200. It is supported when the library knows what it is talking to.

Does LlamaIndex support GPT-6.1 Sol?

Not yet: llama-index-llms-openai 0.8.2 raises ValueError: Unknown model 'gpt-6.1-sol' when you read OpenAI(model="gpt-6.1-sol").metadata, which LlamaIndex reads when it sizes prompts. The same call for gpt-6-sol returns a 1,050,000-token context window. LlamaIndex keeps an explicit table of OpenAI model names, so a new name fails until the table gets a row. Its latest release is from 25 September, four days before GPT-6.1 Sol existed.

Does LangChain support GPT-6.1 Sol?

Not in its model profiles yet. langchain-openai 1.6.6 returns a profile with 1,050,000 input tokens for gpt-6-sol and an empty profile for gpt-6.1-sol. Anything that reads the profile to decide context limits, such as a context-limit check, gets no answer. In JavaScript, @langchain/openai 1.6.0 returns an empty profile for both models, even though it was released on 27 September, five days after GPT-6 Sol launched.

Does LiteLLM know GPT-6.1 Sol’s price?

Yes, but only because LiteLLM downloads its price list at import time. With the default settings, LiteLLM 1.103.1 reports GPT-6.1 Sol at $0.000002 per input token and $0.0000001 per cached token, exactly OpenAI’s published price. The copy bundled inside the package has no entry for GPT-6.1 Sol or GPT-6 Sol, so with LITELLM_LOCAL_MODEL_COST_MAP=True, common in locked-down deployments, both cost $0. That bundled copy shipped at 23:36 UTC on 29 September, after the launch.

A cost dashboard that shows $0 for a new model is not saving you money. It has stopped counting.

Does GPT-6.1 Sol support reasoning effort none?

No. OpenAI’s model page says GPT-6.1 Sol supports low, medium (the default), high, xhigh and max, and that none and minimal are not supported. GPT-6 Sol accepted none. All three libraries that recognise GPT-6.1 Sol already encode this: Pydantic AI's profile rejects none for it, LiteLLM's downloaded list marks it unsupported, and the AI SDK offers none only for GPT-6 Sol and Luna by name. If your code sends none to turn reasoning off, check it before you switch model names.

Think about the last model upgrade you shipped: did you check the library’s model table, or only that the first request came back?

Did the libraries that got GPT-6 Sol wrong fix it?

No. On 26 September we found 5 libraries misreading GPT-6 Sol; on 30 September all 5 still do. crewAI still sizes its context at 6,963 tokens, tiktoken still raises a KeyError, DSPy and the OpenAI Agents SDK still classify it outside the reasoning family, and LangChain.js still returns an empty profile. Two of them, crewAI and LangChain.js, shipped new releases in between. Eight days is longer than most teams wait before switching to a model that is cheaper.

How did we test this?

On 30 September 2026 at 04:56 UTC, about twelve hours after the launch reached Hacker News, we installed the latest PyPI or npm release of each library into its own clean environment and called the function it uses to look up a model: LlamaIndex’s openai_modelname_to_contextsize() and .metadata, LangChain's model profile, LiteLLM's get_model_info() with and without the local map, crewAI's get_context_window_size(), tiktoken's encoding_name_for_model(), DSPy's _is_openai_reasoning_model(), the Agents SDK's gpt_5_reasoning_settings_required(), Pydantic AI's KnownModelName, and the AI SDK's own capability function, executed from its shipped bundle. gpt-6-sol was the control. No API calls were made.

What this does not show

• A request can still succeed. We tested what each library believes about the model, not whether OpenAI accepts the call.
• Twelve hours is a snapshot. Maintainers fix these quickly; we re-ran the probe before publishing, and the status line above has a one-line re-check.
• We did not test every path. LangChain’s empty profile matters only to code that reads it; crewAI’s context figure matters only when a conversation overflows.
• Ten libraries, not the ecosystem. We kept the nine from our 26 September probe and added the Vercel AI SDK. tokencost and js-tiktoken were dropped again because they fail the control: neither knows GPT-6 Sol either.

FAQ

What is the GPT-6.1 Sol model ID? gpt-6.1-sol in the OpenAI API, per OpenAI's model page.

How much does GPT-6.1 Sol cost? $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens at standard rates.

What is GPT-6.1 Sol’s context window? 1,050,000 tokens, with up to 128,000 output tokens.

How do I use GPT-6.1 Sol in LlamaIndex before a fix ships? Until the model table has a row, LlamaIndex’s OpenAILike class with an explicit context_window avoids the name lookup; we did not test it end to end.

When will we re-check? Our follow-up re-probes GPT-6 Sol, GPT-6.1 Sol and Claude Sonnet 5.5 together in the first week of October.

Which library in your stack found out about a new model last?

We publish a measurement a day, each with the method attached. The next framework re-probe counts how many days each of these ten libraries takes to catch up on GPT-6.1 Sol — follow Decoding AI to get it.

If this saved you a debugging session, clap so more developers see it, tell us in the comments which library in your stack learned about a new model last, and follow Decoding AI — then hit subscribe to get the re-probe by email.

Sources & References

• Decoding AI, Is Your AI Framework Ready for GPT-6 Sol? We Probed 9 Libraries — the 26 September baseline this re-probe extends.
• Decoding AI, Does Your AI Framework Support Claude Sonnet 5.5? We Timed 10 Libraries.
• Decoding AI, Switching AI Models in Production Isn’t Routing — why model switches break around the call.
• OpenAI, Introducing GPT-6.1 Sol, 29 September 2026.
• OpenAI API docs, GPT-6.1 Sol model page, read 30 September 2026.
• TechCrunch, OpenAI launches GPT-6.1 Sol, 29 September 2026.
• PyPI and npm registries for each version and upload time, 30 September 2026. Probe scripts and raw output are kept with this post’s draft.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI