Anthropic released Claude Fable 5.2 and Mythos 5.2 at $8/$40 per million input/output tokens, with an optional fast mode at $16/$80, and recommends Opus 5.5 as the daily driver for most work.
How we checked
Anthropic launch page confirms both models, the $8/$40 and $16/$80 prices, and Opus 5.5 as daily driver
Why it matters
New prices and Anthropic's routing advice change which Claude model you default to and when moving up to Fable is worth the cost.
Try this
Keep Opus 5.5 as your default and send only your hardest tasks to Fable 5.2, then compare the cost per finished task.
Google made Gemini 4 Carbon available to paid developers via the Gemini API in AI Studio at $2/$10 per million input/output tokens until November 30, 2026, then $4/$20. Argon stays the recommended default.
How we checked · Why it matters · Try this
How we checked
Google DeepMind launch post states availability, introductory price, the November 30 price footnote, and Argon guidance
Why it matters
A new frontier model at a launch price that doubles in three weeks, so cost tests done now will be out of date by December.
Try this
Run one of your hard multi-step tasks on Carbon in AI Studio and price it at both the launch rate and the December rate.
Gemini 4 Carbon is generally available in GitHub Copilot for Pro+, Max, Business and Enterprise users across VS Code, Copilot CLI, the coding agent and JetBrains, billed at provider list pricing.
How we checked · Why it matters · Try this
How we checked
GitHub changelog entry confirms plans, clients, billing and default enablement
Why it matters
You can try Carbon inside an existing Copilot setup. Admins should know that new models are turned on by default unless the model policy turns them off.
Try this
If you administer Copilot Business or Enterprise, check the model policy in Copilot settings to decide whether Carbon stays enabled.
OpenAI's GPT-Live voice model (gpt-live-1) listens while it speaks and hands work to a backend agent. Delegation runs either through Responses or through your own client harness, and the mode is chosen when the session starts.
How we checked · Why it matters · Try this
How we checked
OpenAI Live guide describes full duplex and both delegation modes; official starter repo targets gpt-live-1
Why it matters
Voice agents no longer have to take strict turns. Your existing agent backend can sit behind a live conversation.
Try this
Read the open-source openai-live-console starter to see both delegation modes before you design a voice front end.
Latent Space's hands-on test of open-weights Naive-N0.5-Flash found a 3.1 s median response vs 5.3 s (GLM 5.3 Flash) and 10.1 s (DeepSeek V4.1 Flash), but far more thinking tokens on hard questions.
How we checked · Why it matters · Try this
How we checked
Hugging Face org lists the model and FP8 drafter; test code repo exists; timings are Latent Space's own
Why it matters
The headline 2,100 tokens per second is a vendor peak figure. Median latency and token use are what you actually pay for.
Try this
Use the open llm-verbosity repo's method to measure median latency and completion tokens on your own prompts before switching models.
OpenAI's GPT-Live prompting guide says to have the model emit <|SpawnThinking|> before any answer that needs backend work, and to use session.commentary.append to give it lines to say while delegated work runs.
How we checked · Why it matters · Try this
How we checked
Live prompting guide confirms the SpawnThinking instruction and the session.commentary.append usage
Why it matters
These are concrete levers against a voice agent that answers before its backend has finished, or goes silent while it waits.
Try this
Add explicit handoff conditions and a <|SpawnThinking|> instruction to your GPT-Live system prompt, then test with requests that need lookups.
Resend serves a machine-readable version of its pricing page when you add .md to the URL, so an agent can read the price tables in one request instead of working a slider.
How we checked · Why it matters · Try this
How we checked
Fetched resend.com/pricing.md; it returns markdown with pricing tables
Why it matters
Coding agents increasingly choose which APIs to use. Pages they can read easily may decide whether your product gets picked.
Try this
Publish a .md version of your pricing and docs pages, then ask an agent a pricing question and see whether it answers correctly.
llm-gemini 0.36 shipped on PyPI on November 10, 2026, the day Gemini 4 Carbon went public. Simon Willison tested Carbon at low, medium and high thinking effort; medium is the default.
How we checked · Why it matters · Try this
How we checked
PyPI JSON shows version 0.36 uploaded 2026-11-10; no sandbox run
Why it matters
You can try Carbon from the command line with the llm tool without writing any API code.
Try this
Run `llm install -U llm-gemini`, then `llm models`, to see which Carbon model IDs the plugin exposes.
Google is previewing /swarm in Antigravity, which turns one prompt into a coordinated team of agents. Each role can use a different model, such as Carbon for architecture and auditing.
How we checked · Why it matters · Try this
How we checked
Google DeepMind launch post describes the /swarm preview; not run
Why it matters
Multi-agent orchestration is now a built-in command in an IDE, not something you have to build yourself.
Try this
In Antigravity, give /swarm a small, well-scoped migration task and check each agent's changes before merging.
Gemini CLI v0.69.0-preview.0 adds a /stats heatmap of daily token usage, standalone theme.json themes, and a way to override the 60 s API header timeout (GEMINI_API_HEADERS_TIMEOUT_MS).
How we checked · Why it matters · Try this
How we checked
Official GitHub release notes list all three changes; preview build, not run
Why it matters
Easier token tracking, plus a fix for long requests that hit the default timeout.
Try this
After upgrading to the preview, run /stats heatmap to see which days use the most tokens.
The Microsoft C++ Language Server for Copilot CLI now bundles its MSBuild compile-commands extractor and runs it automatically, so agents get code intelligence without a separate setup skill.
How we checked · Why it matters · Try this
How we checked
GitHub changelog entry describes the bundled extractor; not run
Why it matters
Less setup and fewer tokens spent before an agent can navigate a large MSBuild C++ codebase.
Try this
On an MSBuild project, install the C++ language server plugin and confirm Copilot CLI generates the compilation database itself.
Anthropic says a team of 1,000 Fable 5.2 agents fully formalized Fermat's Last Theorem in Lean in 30 hours, while Opus 5.5 with the same setup finished 26% in 18 hours.
How we checked · Why it matters
How we checked
Stated on Anthropic's launch page; no independent check of the Lean artifacts
Why it matters
A data point for large agent swarms on long formal tasks, but it comes only from Anthropic.