All issues
The AI-Ready Professional

Issue 016 min read

Gemini 4 Carbon, Hydrogen, and the cost of AI coding

Nine AI takeaways for software professionals, from coding model costs and tool budgets to model routing, agent memory, and review interfaces.

ISO week 46 · November 9–13, 2026

Tool calls and reasoning settings can swing the cost of similar coding results. This week’s releases also give software teams more control over which models handle requests, how agents present decisions, and where they run.

1. Tool definitions and results can dominate agent costs

OpenRouter had Claude Sonnet 5.5 find a GitHub release and three open bugs. The GitHub CLI used about 1,400 input tokens, the CLI with an instruction guide used 3,300, and the GitHub Model Context Protocol (MCP) server with 46 tools loaded used 59,400. With two tools and a cap on response size, the MCP version used about 5,400. All four setups answered correctly. OpenRouter’s GitHub tool comparison

Tool definitions add to the prompt, and a large tool result can be sent back to the model on every subsequent call. In this test, most of the release payload described downloadable files. The lookup needed the tag and date. Returning just the fields a workflow needs reduces that repeated cost. OpenRouter’s breakdown of tool responses

2. Coding model costs diverge even when scores are close

Google opened Gemini 4 Carbon to paying API customers at the same introductory rate as Argon, $2 input and $10 output per million tokens. Gemini 4 Hydrogen began rolling out at $1.50 input and $7.50 output. Anthropic released Claude Fable 5.2 at $8 input and $40 output in standard mode. Google’s Carbon launch · Google’s Hydrogen launch · Anthropic’s Fable 5.2 launch

These are selected CursorBench 4.0 results for multi-file tasks drawn from real Cursor sessions. The cost is the average per task at the listed reasoning setting. CursorBench 4.0 results and methodology

Model and effort Score Cost per task
Gemini 4 Carbon · High 60.4% $9.22
Claude Fable 5.2 · Max 60.0% $32.79
Gemini 4 Hydrogen · Extra High 57.8% $5.86
Gemini 4 Hydrogen · High 54.3% $5.59
Gemini 4 Argon · High 52.9% $4.57

Carbon High and Fable Max scored within half a point of each other, with Carbon costing about 72% less per task. Hydrogen Extra High gained 3.5 points over Hydrogen High for another $0.27 per task. Cursor notes that small differences can fall within normal variation. CursorBench 4.0 results and methodology

For recurring work like extracting fields from support tickets, Meta’s Muse Ember launched at $0.10 input and $0.50 output per million tokens. Meta reports that Ember scored ahead of Claude Haiku 5.5 on Terminal-Bench 4.0 and behind it on FrontierCode 2.0, which grades whether code changes are ready to merge. That distinction matters for teams choosing between a model for repeatable tool calls and one for producing reviewable patches. Meta’s Muse Ember launch and benchmarks

3. Model routing can stay within an approved provider list

OpenRouter’s Auto Code selects a coding model for each request using signals such as task difficulty, pricing, provider health, and cache status. It can route within an approved set of models and providers and respect zero-data-retention requirements. The API can return routing metadata, so a team can see whether a change in model or provider coincided with a shift in cost or latency. OpenRouter’s Auto and Auto Code release

4. Classification APIs can return confidence scores

The Gemini Decisions API scores predefined choices, yes/no checks, and rubric items against the same input. It accepts text, images, audio, video, and PDFs, and can score up to 1,000 independent questions in one request. Google lists $0.10 per million input tokens with zero output tokens billed. In ticket triage, an app could route high-confidence labels to the right queue and send ambiguous cases for review. A threshold derived from labeled tickets gives the team a way to balance misroutes against review volume. Google’s Gemini Decisions announcement

5. Agent monitoring catches some harmful actions after execution

Google evaluated Agent Anomaly Detection on five public benchmarks created by other teams. It flagged 87% of harmful instructions that the agent acted on for review after execution. False alerts ranged from 2% to 16% of benign sessions, depending on the tools involved. It caught just 4 of 18 executed cases involving poisoned long-term memory. For actions like changing production settings, scoped credentials and approval checks can stop an unauthorized write before monitoring would see it. Google’s benchmark report

6. Agent memory is easier to correct when notes have a source

Anthropic describes a memory design with searchable session transcripts, notes written by the agent, and a small amount of context loaded at the start of a new session. Between sessions, the agent can compare its notes with the original transcript and correct mistakes. In a long-running debugging session, that gives an agent a way to trace a note such as “the timeout is in the queue worker” back to the log or test result that supports it. Anthropic’s research on agent memory

7. Agent review screens can use an app’s own components

Google’s A2UI 1.0 lets an agent describe a view that the host app renders from a catalog of registered components. The protocol supports interactions such as picking an option or checking availability, with renderers for React, Angular, and other platforms. In Jira, Atlassian is using it to put interactive follow-up cards directly into work items. The same pattern fits a support workflow where a reviewer can edit an agent’s proposed owner and priority before the app saves the change. Google’s A2UI 1.0 announcement · A2UI documentation

8. Parallel coding agents work best with clear ownership

Google Antigravity’s Swarm preview gives each worker a defined assignment and uses a shared board for coordination. Workers share a workspace by default and can get separate working copies for concurrent edits. A verifier can read diffs and run tests. A change split across independent packages fits this setup because the workers can make progress without editing the same files. Each worker is a separate agent, so the setup uses quota faster than a single conversation. Antigravity Swarm documentation

9. Self-hosting the agent service lets teams control its data and runtime

Anthropic’s self-hosted Claude Managed Agents private beta puts the agent loop, execution sandboxes, credentials, and session database in an organization’s own cloud or on-premises environment. The organization operates the service and chooses where model requests go, such as a supported cloud provider or its own gateway. That gives infrastructure teams control over where session history is stored and how credentials are handled, with the operational responsibility of running the service and its workers. Anthropic’s self-hosting announcement