GPT-6.1 Sol replaces GPT-6 Sol after just 7 days, with near-Astra intelligence

TL;DR Highlight

GPT-6.1 Sol scores 1 point below Astra at under a quarter of the cost per task

Who Should Read

A backend developer running agent or coding workloads on GPT models who is weighing a model swap and cost optimization

Core Mechanics

Only 1 point behind Astra
- It gains 4 points over GPT-6 Sol and 5 over GPT-5.6 Sol on the Intelligence Index.
- That puts it 1 point below GPT-6 Astra.
Same price, bigger cache discount
- Pricing matches GPT-6 Sol at $2 input / $10 output per million tokens.
- The cache read discount rises from 90% to 95%, so blended agentic pricing is slightly lower.
Cost per task dropped sharply
- At max effort it costs $0.72 per Intelligence Index task, under a quarter of GPT-6 Astra ($3.26).
- That is 31% less than GPT-6 Sol ($1.05) and 64% less than GPT-5.6 Sol ($1.99).
Agentic work and hallucination improved
- Terminal-Bench 4.0 is up 12 points and Humanity's Last Exam is up 5.
- AA-Omniscience Accuracy is up 8 points and hallucination rate fell from 60% to 54%.
It uses 10-30% more output tokens

Community Reactions

Usage runs out faster, some say
- One commenter said they haven't noticed performance jumps since GPT-5.5, only that their usage ends sooner.
- Another countered that people say this every release, yet AI has changed dramatically over two years.
Mixed hands-on reports
- One commenter called 6.1 nerfed and unusable.
- Another said ChatGPT has been a mess since launch.
Request for per-effort comparisons
- A Claude user wanted the benchmark site to show several thinking modes of a single model.
- That would help them gauge what high or medium effort actually gives them.
Reactions to pricing
- One commenter said prices are racing to the bottom.
- Others speculated the 7-day replacement was a reaction to Anthropic's launch.

How to Apply

Build a cost comparison table today
- Pick 5-10 representative tasks from your service and run them on GPT-6 Sol and GPT-6.1 Sol.
- Record accuracy, input/output tokens, and cost per task in one table.
Try 6.1 Sol where you call GPT-6 Astra
- Start with cost-sensitive paths and A/B test them on 6.1 Sol.
- Use the source's baseline: 1 point lower intelligence at under a quarter of the cost per task.
Run coding tasks at both xhigh and max
- Execute the same coding tasks at xhigh and max effort.
- Compare score and cost to set your default effort.
Audit your prompt cache layout
- With a 95% cache read discount, pin the system prompt and tool definitions at the front.
- Put per-request content at the end to raise cache hit rate.

Try It

copy & run

# Paste into Claude or ChatGPT (model swap evaluation planning prompt)

You are an LLM cost optimization consultant. We are considering moving from GPT-6 Sol to GPT-6.1 Sol.
Known facts: pricing is the same at $2 input / $10 output per million tokens, the cache read discount rises from 90% to 95%, and output tokens run about 10-30% higher.

Our service:
- Main workload: {{workload description}}
- Monthly input tokens: {{input tokens}}
- Monthly output tokens: {{output tokens}}
- Cache hit rate: {{cache hit rate}}

Instructions:
1. Calculate monthly cost before and after the swap, showing the formulas.
2. Compare two cases: output tokens up 10% and up 30%.
3. Suggest 5 A/B test items to run before switching.

Next step: rerun the calculation with token counts from your real logs

Terminology

Intelligence IndexA composite score from Artificial Analysis that combines several evaluations. Think of it as a model's overall report card.

Cost per TaskThe average cost to complete one evaluation task. It reflects the actual bill rather than the unit price.

Pareto frontierThe edge of models where no alternative is better on both cost and performance. It means no cheaper model exists at that intelligence.

effortA setting for how long the model thinks, with levels like low, medium, xhigh and max.

Cache read discountThe discount applied when re-reading already cached input. The more of your prompt prefix you reuse, the cheaper it gets.