assafelovic/gpt-researcher · error · ValueError

max_tokens={max_tokens} exceeds the largest output limit of

Error message

max_tokens={max_tokens} exceeds the largest output limit of any currently available model (128k as of late 2025). Check your FAST_TOKEN_LIMIT / SMART_TOKEN_LIMIT / STRATEGIC_TOKEN_LIMIT env vars for typos.

What it means

A sanity guard in create_chat_completion: any max_tokens above 200,000 is rejected because no current model supports more than ~128k output tokens. Values this large almost always come from a typo'd FAST_TOKEN_LIMIT / SMART_TOKEN_LIMIT / STRATEGIC_TOKEN_LIMIT env var (e.g. an extra digit) rather than a genuine need.

Source

Thrown at gpt_researcher/utils/llm.py:76

        temperature (float, optional): The temperature to use. Defaults to 0.4.
        max_tokens (int, optional): The max tokens to use. Defaults to 4000.
        llm_provider (str, optional): The LLM Provider to use.
        stream (bool): Whether to stream the response. Defaults to False.
        webocket (WebSocket): The websocket used in the currect request,
        llm_kwargs (dict[str, Any], optional): Additional LLM keyword arguments. Defaults to None.
        cost_callback: Callback function for updating cost.
        reasoning_effort (str, optional): Reasoning effort for OpenAI's reasoning models. Defaults to 'low'.
        **kwargs: Additional keyword arguments.
    Returns:
        str: The response from the chat completion.
    """
    # validate input
    if model is None:
        raise ValueError("Model cannot be None")
    # Sanity guard against absurd values (e.g., env var typos). The actual
    # per-model output limits are enforced by the upstream provider.
    if max_tokens is not None and max_tokens > 200_000:
        raise ValueError(
            f"max_tokens={max_tokens} exceeds the largest output limit of "
            "any currently available model (128k as of late 2025). "
            "Check your FAST_TOKEN_LIMIT / SMART_TOKEN_LIMIT / "
            "STRATEGIC_TOKEN_LIMIT env vars for typos."
        )

    # Get the provider from supported providers
    provider_kwargs = {'model': model}

    if llm_kwargs:
        provider_kwargs.update(llm_kwargs)
    elif os.environ.get("LLM_KWARGS"):
        import json
        try:
            provider_kwargs.update(json.loads(os.environ["LLM_KWARGS"]))
        except json.JSONDecodeError:
            pass

View on GitHub (pinned to 6f998577d5)

Solutions

1. Fix the env var: check FAST_TOKEN_LIMIT, SMART_TOKEN_LIMIT, STRATEGIC_TOKEN_LIMIT for extra digits and set realistic values (e.g. 4000–128000)
2. If calling directly, pass max_tokens=None to use provider defaults or a value ≤128000
3. Print your token-limit settings before the call to confirm what's actually being sent

Example fix

# before
export FAST_TOKEN_LIMIT=4000000  # ValueError: max_tokens=4000000 exceeds...

# after
export FAST_TOKEN_LIMIT=400000

MAX_OUTPUT_TOKENS = 200_000
limit = int(os.getenv('FAST_TOKEN_LIMIT', 400000))
assert limit <= MAX_OUTPUT_TOKENS, f'FAST_TOKEN_LIMIT={limit} is invalid'

def valid_token_limit(v) -> bool:
    return v is None or (isinstance(v, int) and 0 < v <= 200_000)

try:
    resp = await create_chat_completion(prompt, model, max_tokens=limit)
except ValueError as e:
    if 'max_tokens' in str(e):
        resp = await create_chat_completion(prompt, model)  # provider default
    else:
        raise

When it happens

Trigger: Passing max_tokens > 200000, usually because a *_TOKEN_LIMIT env var was set incorrectly (e.g. 4000000 instead of 400000) or a config default got corrupted; the value is forwarded from token_limit settings into create_chat_completion.

Common situations: Copy-pasting token limits between projects with an extra digit; confusing context-window size (1M+) with output-token limits; env var typos after an upgrade that introduced these variables.

Related errors

• rate_limit_delay must be a number, got {rate_limit_delay!r}
• Invalid reasoning effort: {reasoning_effort_str}. Valid opti
• Model cannot be None
• rate_limit_delay must be non-negative, got {delay}
• LLM_PROVIDER, FAST_LLM_MODEL and SMART_LLM_MODEL are depreca

AI-assisted analysis of assafelovic/gpt-researcher@6f998577d5 (2026-08-28). Data as JSON: /api/errors/063eac0fc5445e52. Report an issue: GitHub.