Model Watch: The Subtle but Important Steps Toward Autonomous Agents
Notes by SDG ·
A turning point: models that assemble work
Last week we looked at how models began to execute code in‑loop.
Since then, the frontier labs have continued to blur the line between language and action.
Large‑language‑models are no longer just chatbots; they’re becoming systems that plan, reason, execute and verify.
Below are the major developments from the last two weeks of November.
🟢 Anthropic:
Claude Opus 4.5 raises the bar for engineering
- Opus 4.5 released (Nov 24) – Anthropic launched its new flagship model, Claude Opus 4.5, which improves reasoning, math and coding while cutting prices. It retains a 200 k token context and adds an “effort” parameter to trade off speed versus depth. The model outscored previous versions on Anthropic’s internal engineering assessment and can now handle tasks that previously took two hours in ~30 minutes . Opus 4.5 uses up to 65 % fewer tokens on long‑horizon coding tasks and is priced at $5/M input tokens and $25/M output tokens —significantly undercutting GPT‑5.1 and Gemini 3.
- Improved tool‑use and alignment – The model inherits the tool‑use improvements from Sonnet 4.5 and introduces an effort parameter (“low”, “medium”, “high”) to adjust computational work . Combined with Anthropic’s zoom feature (which lets Claude control your computer) and the ability to preserve thinking blocks across turns , Opus 4.5 is marketed as a collaborator rather than just a text generator.
🔵 OpenAI:
GPT‑5.1 and Codex‑Max push adaptive reasoning
- GPT‑5.1 for developers – Released Nov 13, GPT‑5.1 dynamically adjusts how much “thinking” it does based on task complexity . A new “no reasoning” mode lets it skip unnecessary reasoning for quick replies . Extended prompt caching (up to 24 hours) improves follow‑up response time . OpenAI also introduced an apply_patch tool for reliable code editing and a shell tool to run shell commands .
- GPT‑5.1 Codex‑Max – On Nov 19, OpenAI unveiled Codex‑Max, an agentic coding model designed for project‑scale tasks. It replaces GPT‑5.1 Codex across Codex surfaces and excels at long‑horizon tasks thanks to a compaction mechanism that retains relevant context while discarding irrelevant details . Codex‑Max can complete multi‑day coding sessions and uses ~30 % fewer tokens than its predecessor . Benchmarks show it achieving 77.9 % accuracy on SWE‑Bench Verified, edging past Gemini 3 Pro .
- New ChatGPT features – OpenAI released IndQA, a benchmark testing models on Indian languages and culture; this will push models toward better cultural reasoning. It added ChatGPT Apps with Peloton and TripAdvisor, allowing ChatGPT to deliver fitness coaching and travel planning through trusted partners. A new interrupt & refine feature lets users steer long tasks mid‑execution, improving research and agentic workflows.
🟣 Google:
Gemini 3 Pro & Deep Think arrive; Antigravity goes live
- Gemini 3 Pro – Google’s most intelligent model to date unifies reasoning, coding and multimodal understanding. It outperforms Gemini 2.5 Pro on every major benchmark, topping the LMArena leaderboard with a 1501 Elo and achieving high scores on Humanity’s Last Exam (37.5 %) and GPQA Diamond (91.9 %) . Gemini 3 offers state‑of‑the‑art scores in math and multimodal tasks such as MMMU‑Pro and Video‑MMMU .
- Gemini 3 Deep Think – An enhanced reasoning mode that will be released to safety testers first, Deep Think further boosts performance on complex benchmarks (e.g., 41 % on Humanity’s Last Exam and 93.8 % on GPQA Diamond) . Gemini 3 is available across Search (AI Mode), the Gemini app, Vertex AI, and Google’s new developer platform .
- Antigravity platform – Google launched Antigravity, an agentic development environment where Gemini 3‑powered agents plan, code and document their own work. The platform groups low‑level tool calls into human‑readable artifacts and provides editor and manager surfaces for orchestrating multiple agents in parallel. It emphasises trust and feedback loops .
🟡 xAI: Grok 4.1 & 4.1 Fast show 2M‑token context & agent tools
- Grok 4.1 – xAI’s update emphasises creativity and emotional nuance. Grok 4.1 was rolled out gradually between Nov 1–14, and user preference tests show it is preferred ~65 % of the time over Grok 4 . It also scores highly on style‑control and emotional intelligence benchmarks .
- Grok 4.1 Fast & Agent Tools API (Nov 19) – xAI launched Grok 4.1 Fast, a tool‑calling model with a 2 million‑token context window designed for long‑horizon agentic tasks . Paired with the Agent Tools API, Grok 4.1 Fast can browse the web, search X posts, execute code, and retrieve documents, enabling fully autonomous agents . The API handles API keys, rate limits and retrieval pipelines so developers can focus on high‑level agent logic . Grok 4.1 Fast is trained via reinforcement learning on diverse tool use, achieving 100 % on the τ²‑bench Telecom benchmark and 72 % on Berkeley Function‑Calling v4 . It excels across long‑context, multi‑turn scenarios .
🪐 Other notable updates
- Claude Sonnet 4.5 & Haiku 4.5 (Anthropic) – These mid‑sized and small models, released earlier this fall, inherited the tool‑use and memory improvements of Opus 4.5. They handle long‑context coding tasks more efficiently and offer lower pricing for scaled deployments .
- Olmo 3 (AI2) – The Allen Institute’s open‑source LLM includes base models (7B and 32B) with Think variants that excel at programming, reading comprehension, math and long‑context tasks up to 65 k tokens; the suite launched Nov 20 and is freely downloadable .
- P1 physics models – Shanghai AI Lab introduced P1‑235B‑A22B, an open‑source model trained via reinforcement learning that achieved gold medal performance on the International Physics Olympiad 2025 and outperformed previous open‑source models on math and coding tasks.
🔚 Appendix – Anthropic’s MCP upgrade
(example)
Anthropic’s MCP upgrade allows agents to move from sequential tool calls to code‑driven loops executed in a sandboxed environment. Here’s a practical example that illustrates the difference.
Scenario-
Goal: Which product category grew fastest last month?
Pre-MCP Workflow
The agent reads a CSV file and then calls tools sequentially: read_csv, filter_by_date, group_by, sum, sort. Each step requires restating parameters and passing intermediate results, making the workflow slow and token heavy.
Post MCP Workflow
Using MCP, the agent generates a block of Python code, sends it to the MCP server for execution, and receives structured results. The code can group, sum and sort the data in one go and preserve state. The agent iterates by modifying the code rather than re-invoking multiple tools.
In the post‑MCP setup, typical workflows consume ≈98 % fewer tokens and run up to 10× faster because the agent doesn’t need to re‑describe tools or carry context back and forth. Errors are confined to code tracebacks rather than derailing entire prompt chains.
Summary
The last two weeks saw frontier models transition from monolithic text engines to modular agents capable of executing code, adapting reasoning, and grounding decisions in data.
These advances hint at a future where language models serve as orchestrators for entire workflows—planning, executing and verifying work autonomously—while safety research moves toward provable guarantees.