GPT-6.1 Sol vs Claude Fable 5.1: 10x the Cost, Same Score?

GPT-6.1 Sol costs a fifth of Claude Fable 5.1 per token and scored level with it on the one coding board that lists both. On Terminal-Bench 4.0, tbench.ai has Sol at 58.2% (run in Codex) and Fable 5.1 at 57.9% (run in Claude Code). That gap sits inside the 95% confidence intervals, so neither model leads. The full run cost $0.6k for Sol and $6.2k for Fable 5.1. Fable 5.1 scores higher on some broader indexes, so route routine work to Sol and test Fable 5.1 where Sol falls short. OpenAI’s price-equal tier to Fable 5.1 is GPT-6 Astra. All data is as of October 19, 2026.

GPT-6.1 Sol vs Claude Fable 5.1: price and specs

Sol lists at one fifth of Fable 5.1’s price on input and output tokens, and the context and output limits are close. The numbers below come from OpenAI’s GPT-6.1 Sol page and Anthropic’s pricing page.

Two details change a bill. Cache reads are only 2.5x apart ($0.10 against $0.25), so a cache-heavy loop narrows the gap. And OpenAI prices Sol’s Fast mode at 2x Standard and Ultrafast at 6x, so a Sol call tuned for speed no longer costs a fifth of Fable 5.1.

Where GPT-6 Astra fits

Astra is the OpenAI model that sits at Fable 5.1’s price: $10 input and $50 output per 1M tokens on OpenAI’s pricing page. OpenAI calls Astra its highest-intelligence tier and describes Sol as near-Astra performance at a lower cost. On tbench.ai, Astra scores 58.2% with a 2.8-point margin at a $3.3k run cost, level with both models here. If you want OpenAI’s top tier instead of its cost-efficient one, compare GPT-6 Astra with Fable 5.1. The rest of this page stays on Sol.

Sol vs Fable benchmarks: where each model leads

On the benchmark where both models share a table, they tie. On broader indexes Fable 5.1 leads, and Sol leads a few agent and science tasks. The table below keeps each number with its owner, because the sources do not agree and should not be averaged.

The tbench.ai runs use a different agent harness for each model, Codex for Sol and Claude Code for Fable 5.1, so part of the score belongs to the harness. The Fable 5.1 rows on Artificial Analysis are labeled as run with default fallback, and Vals AI ran Fable 5.1 with Claude Opus 5 and Claude Opus 4.8 as fallbacks, with a 2.10% fallback rate. Fable 5.1’s own Terminal-Bench 4.0 score runs from 52.0% to 58.08% across Artificial Analysis, Vals AI, tbench.ai and Anthropic’s launch table (55.8%), depending on effort and harness. Anthropic’s launch table also compares Fable 5.1 against GPT-5.6 Sol, not 6.1 Sol.

For vendor claims, OpenAI reports 75.2% for Sol on DeepSWE v1.1 at high reasoning effort, up from 68.8% for GPT-6 Sol at max, in its developer community announcement. That compares Sol with OpenAI’s own earlier model, not with Fable 5.1. Neither vendor has published SWE-Bench Pro for the new pair, and Scale’s public SWE-Bench Pro leaderboard lists neither model, so any SWE-Bench Pro number for Sol or Fable 5.1 you see is not from the benchmark owner.

Speed is mixed. Artificial Analysis measures 68 output tokens per second for Fable 5.1 and 60 for Sol at max effort, yet Vals AI lists a latency of 76 minutes 52 seconds for Fable 5.1 and 43 minutes 10 seconds for Sol on the Vals Index. If the symptom you see is slow responses or provider queueing, LLM API latency metrics beyond benchmark scores covers what to measure.

Sol vs Fable cost: a worked example

At list price, Sol is 5x cheaper on a 100K-token prompt and about 2.6x cheaper on a 300K-token prompt, because Sol reprices the whole request once input passes 272K tokens. The table uses one request shape: 10,000 output tokens, no caching, no batch, with input at 100K and then 300K. A coding agent that re-reads a 100K-token repo summary on every step looks like the first row. A retrieval job that stuffs 300K tokens of documents into one call looks like the second.

Without the long-context rule, the 300K request on Sol would cost $0.60 plus $0.10, or $0.70. The surcharge adds $0.65, and Fable 5.1 has no equivalent. These are illustrative calculations from list prices, not measured bills.

Per-token price is only half the cost, because Fable 5.1 also spends more tokens on the same work. In tbench.ai’s Terminal-Bench 4.0 run, Sol used 1.5B tokens for $0.6k and Fable 5.1 used 2.7B tokens for $6.2k, so the 5x token price becomes roughly 10x per run. Artificial Analysis shows the same shape per task: $0.72 for Sol at max effort against $7.63 for Fable 5.1, and $0.13 against $2.37 at low effort. Compare cost per solved task on your own prompts, not the rate card.

API differences that break code

Swapping one model ID for the other can fail on tool calling and reasoning settings, because the two vendors enforce different rules. These are the vendors’ own rules from OpenAI’s model page and Anthropic’s Fable 5.1 notes.

GPT-6.1 Sol

• Tool calling needs the Responses API. Chat Completions works only without tools.
• reasoning.effort accepts low, medium (the default), high, xhigh and max. none and minimal are not supported, so code written for GPT-5.6 Sol’s none setting needs a new default.

Claude Fable 5.1

• A forced tool_choice of {"type": "any"} or {"type": "tool", ...} returns a 400 invalid_request_error. Keep auto and use strict tool use or structured outputs for schema-valid JSON.
• Adaptive thinking is always on. Disabling it, prefilling the assistant reply, or sending a non-default temperature, top_p or top_k all return a 400.
• Only Fable 5.1 and Mythos 5.1 can read Fable 5.1’s thinking blocks, and editing earlier turns of a conversation invalidates them. Keep the message history append-only.
• The classifiers can end a request with stop_reason: "refusal". Since September 24, 2026, a refusal that arrives before any output is billed in some categories.
• Anthropic keeps Fable 5.1 requests for 30 days and does not offer zero data retention unless it authorizes it.

If you reach either model through MixRoute, send a tool-calling request down the exact route you plan to use before you move production traffic.

Who can use each model today

Both models are generally available, in different places. GPT-6.1 Sol is in the OpenAI API, Codex and ChatGPT Work, and it is not in ChatGPT Chat, according to OpenAI’s Codex models page. Claude Fable 5.1 is available to all customers of the Claude API and on Amazon Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS. Do not confuse it with Claude Mythos 5.1, which is the same model with looser safeguards and is open only through Anthropic’s trusted-access programs. Anthropic’s own model overview says to start with Claude Opus 5.5 for most workloads and use Fable 5.1 when evals at higher effort still fall short.

Which to pick, and how to run both

Run routine, high-volume work on Sol and keep Fable 5.1 for the jobs where Sol’s answers fall short. The numbers above support that split. Sol costs about a tenth per run on the shared Terminal-Bench 4.0 table with an equal score, and it is ahead on Vals’ Terminal-Bench-Science and scores higher on AutomationBench-AA. Fable 5.1 is ahead on the Vals Index and scores higher on GDPval-AA, Humanity’s Last Exam and SciCode (63.1% against 54.2% on Artificial Analysis).

MixRoute lets you do this with one key. Both models sit behind the same OpenAI-compatible base URL, and you switch by changing the model ID. MixRoute’s GPT-6.1 Sol and Claude Fable 5.1 model pages list $2 and $10, and $10 and $50 per 1M tokens, including Sol’s $4 and $15 tier above 272K tokens. Fable 5.1 is also reachable through MixRoute’s Anthropic-compatible /v1/messages endpoint.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_MIXROUTE_KEY",
    base_url="https://api.mixroute.ai/v1",
)

def ask(model, prompt):
    resp = client.chat.completions.create(
        model=model,
        messages=[{"role": "user", "content": prompt}],
    )
    return resp.choices[0].message.content

# Routine work on Sol
ask("gpt-6.1-sol", "Summarize this stack trace and name the likely cause: ...")

# Escalate the hard case to Fable 5.1
ask("claude-fable-5-1", "Review this migration plan for ordering bugs: ...")

If you would rather not write the routing rule yourself, Smart Routing picks a model for each request by complexity and cost. It adds no routing fee, you pay only for the model it selects, and it adds one triage step before the request runs. For another model comparison, see the GLM-5.3 vs Opus 5.5 vs Fable 5.1 comparison.

What changed since GPT-5.6 Sol vs Claude Fable 5

Both models were replaced, and Sol now costs half as much and is open to API users. If you arrived from the earlier GPT-5.6 Sol vs Claude Fable 5 comparison, this table shows what moved.

The earlier article compared launch-chart scores on Terminal-Bench 2.1 and SWE-Bench Pro. tbench.ai now ranks both generations on Terminal-Bench 4.0, which is why this page uses it, and SWE-Bench Pro is unpublished for the new pair.

FAQ

Is GPT-6.1 Sol better than Claude Fable 5.1? It depends on the job. On Terminal-Bench 4.0 they tie at 58.2% and 57.9% on tbench.ai. Fable 5.1 leads the Vals Index (65.83% against 61.15%), while Sol leads Terminal-Bench-Science on Vals (52.86% against 40.00%) and costs a fifth as much per token. Score alone is the wrong number to compare, so divide cost by tasks solved on your own prompts.

How much cheaper is GPT-6.1 Sol than Claude Fable 5.1? At list price Sol is 5x cheaper on input and output tokens: $2 and $10 against $10 and $50 per 1M. Batch discounts both by half, so the ratio holds. Cache reads narrow it to 2.5x ($0.10 against $0.25), and prompts above 272K input tokens narrow it to about 2.6x on a 300K-in, 10K-out request. On tbench.ai’s full run the gap was roughly 10x.

Is GPT-6.1 Sol the same as GPT-6 Astra? No. Astra is OpenAI’s highest-intelligence tier at $10 input and $50 output per 1M tokens, the same list price as Fable 5.1. Sol is the cheaper tier below it, which OpenAI describes as near-Astra performance at lower cost. On tbench.ai both score 58.2% on Terminal-Bench 4.0, with Astra’s run costing $3.3k against $0.6k for Sol.

Can I use GPT-6.1 Sol in ChatGPT? Only in ChatGPT Work and Codex. OpenAI says Sol is not available in ChatGPT Chat, and its Codex page adds that you select it in the model picker when it is available to your account and client. For your own applications, call it through the OpenAI API with the model ID gpt-6.1-sol, using the Responses API if the request includes tools.

Which is better for coding, Sol or Fable? Neither has a proven lead on terminal coding, because the tbench.ai intervals overlap. Neither vendor has published SWE-Bench Pro for these models. Sol finishes the same benchmark at about a tenth of the cost, so start there for volume. tbench.ai also lists Claude Opus 5.5 at 64.8% (+/-3.1) with Claude Code, which is a third model worth testing against both.

How do I call both models through one API? Create one MixRoute key, point an OpenAI-compatible client at https://api.mixroute.ai/v1, and set the model to gpt-6.1-sol or claude-fable-5-1. Nothing else in the client changes. Anthropic SDK users can instead set the base URL to https://api.mixroute.ai and keep the Messages API for Fable 5.1.