One prompt can now drive more of your work More support for defending infrastructure and open source See Claude’s rules updated for newer risks GPT-6 Intelligent UI makes conversations visual and interactive Claude Haiku 5.5 delivers low-cost, high-performance AI Create visual, interactive answers from a simple chat Run high-volume tasks with a cheaper fast model Share new mathematical results on GitHub to accelerate research Easier access to advanced Claude models for security work Run image, audio, and video search on-device with one model Atlassian integration makes company knowledge easier to use Decisions beta speeds up typed answers from text and images Anthropic expands safer access to advanced cyber features Enable text watermarking via API for EU compliance Claude training becomes easier for enterprise teams Anthropic invests in workforce training for enterprise adoption Enterprise adoption and training get easier Google's Gemini 4 Argon makes heavy tasks easier to offload Gemini 4 Argon is built for long professional tasks Use Astra-level performance affordably in daily work One prompt can now drive more of your work More support for defending infrastructure and open source See Claude’s rules updated for newer risks GPT-6 Intelligent UI makes conversations visual and interactive Claude Haiku 5.5 delivers low-cost, high-performance AI Create visual, interactive answers from a simple chat Run high-volume tasks with a cheaper fast model Share new mathematical results on GitHub to accelerate research Easier access to advanced Claude models for security work Run image, audio, and video search on-device with one model Atlassian integration makes company knowledge easier to use Decisions beta speeds up typed answers from text and images Anthropic expands safer access to advanced cyber features Enable text watermarking via API for EU compliance Claude training becomes easier for enterprise teams Anthropic invests in workforce training for enterprise adoption Enterprise adoption and training get easier Google's Gemini 4 Argon makes heavy tasks easier to offload Gemini 4 Argon is built for long professional tasks Use Astra-level performance affordably in daily work

Official sources only. Rumors, leaks, and get-rich schemes are excluded.

← Back to top

AI BriefingOpenAIFeature Updates23:57

AI summarized from verified sources

Responses API with reasoning retention and compaction triples benchmark scores

API developers can efficiently boost model performance with Responses API and reasoning retention settings.

SOURCE CHECK

2 sources

VERIFIED

Sources

Primary / openai.com

Official Blog

Supporting / x.com

Official Blog

Key Points

1Responses API enables retained reasoning
2Context compaction improves token efficiency
3Demonstrated 188% ARC-AGI-3 score improvement

OpenAI applied the Responses API with retained reasoning and context compaction to the ARC-AGI-3 benchmark. GPT-5.6 Sol's score rose 188% while using 6x fewer output tokens. The standard harness discarded reasoning after each move and dropped earlier actions as context filled, forcing restarts. Developers can now adopt these settings to improve real-world performance.

What happened

OpenAI switched the ARC-AGI-3 harness to the Responses API and enabled retained reasoning plus context compaction. This led to a major score improvement for GPT-5.6 Sol.

Impact

API users adopting these settings can expect better performance on long-running agent tasks and reduced token usage. It also highlights that benchmark results depend heavily on settings.