Meta launches Muse Wonder for agents; GPT-6.1 Sol's default context jumps to 500k

Meta Superintelligence Labs released two closed-weight models. The main features are Active Memory, which carries user context across conversations and pulls in what matters, and Active Focus, which keeps working toward a hard goal by trying new approaches while checking in before consequential actions. Standard Muse Wonder is available through the Meta Model API (muse-wonder-1.0) and the Muse Code coding tool. The larger Muse Wonder Full is for Muse Code subscribers only. It takes text, images, audio and video, with a roughly 1M-token context window.

At $4 per million input tokens and $20 per million output, it's a new option priced alongside the top tier. On the Artificial Analysis Intelligence Index, Wonder scores 52 and Wonder Full 56Independent, just behind Claude Opus 5.5 at 58Independent. Meta says Wonder Full leads the long-horizon SWE-Marathon benchmark at 57.5Vendor claim. Teams running long agent jobs should put it on their eval shortlist.

On day 7 of its "28 days of improvements" run, OpenAI raised GPT-6.1 Sol's default context (how much input the model holds at once) in Codex and ChatGPT Work from 272k to 500k tokens, roughly an 80% increase. Users can still set it anywhere up to 1M tokens. When usage comes from a subscription, long-context fees above 272k are waived too.

OpenAI says long tasks should see 2-3x fewer compactionsVendor claim, the step where the conversation gets summarized to fit. Agents working across big codebases should get interrupted less and lose less nuance. It follows the Ultrafast mode we covered earlier. This time the change is in how much the model remembers rather than how fast it answers.

MASS (Multi-Agent Self-Supervision) has a single model play every role: a team of subagents that solve the task, an optimizer that rewrites the team's workflow, and an evaluator that grades the results. Once it finds a better workflow, the model is fine-tuned on the best team's runs, which folds the team's experience back into its weights. Experiments used the open-weight Qwen3.6-27B, and the code is public.

After two rounds, the win rate against the base model on held-out research tasks rose to 69.9%Vendor claim, and score per output token was 1.2-1.6x higher across four benchmarksVendor claim. Multi-agent traces also beat single-agent traces with less data. That's useful for any team thinking about recycling its own agent logs as training data.

Mistral's largest model yet is a mixture-of-experts (MoE) model, an architecture that only activates a subset of expert layers per token. It has 1 trillion parameters in total, 52 billion active. It understands images and documents alongside text and supports more than 160 languages. For now it's a public API preview in Mistral Studio, with weights promised by the end of the month.

At $1.36 per million input tokens and $4.18 per million output, it costs far less than the top closed models, and once weights ship, on-prem and private-cloud deployment become possible. Mistral is focusing on agent and security work, citing 61.7% on DeepSWE v1.1Vendor claim and 93% on CybenchVendor claim. That matters most to European, public-sector and financial teams that care about data sovereignty.