AI Dev Weekly #24: Gemini 3.8 Flash Goes GA, Fable 5.1 Cuts Agent Cache Costs, Agent Plugins 1.0 Ships
AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.
The common thread this week is production readiness. Google gave Gemini 3.8 Flash a stable endpoint and a dated price increase. Anthropic made Fable 5.1 generally available while cutting the cost of repeatedly reading cached context. VS Code shipped a portable plugin format alongside its persistent Agent Host architecture. And GitHub extended content exclusions to the Copilot app and CLI, closing an important policy gap as agents move beyond the IDE.
1. Gemini 3.8 Flash goes GA, but the launch price expires in four months
Google released Gemini 3.8 Flash on September 2 as a generally available production model. The stable API ID is gemini-3.8-flash, with a 1,048,576-token input limit, up to 65,536 output tokens, and low, medium, or high thinking levels. Medium is the default.
The official Gemini 3.8 Flash guide makes two details especially important for production planning:
- introductory pricing is $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026;
- standard pricing doubles to $1.50 input and $7.50 output on January 1, 2027.
Output billing includes thinking tokens, so a higher reasoning level can raise both latency and cost. Google also warns that 3.8 Flash deliberately takes more small reasoning steps and makes more tool calls on difficult tasks. The cheaper launch rate does not guarantee that a long agent run costs less than it did on an earlier Flash model.
The model supports function calling, structured output, caching, code execution, search grounding, file search, Batch, Flex, and Priority inference. Computer use remains in preview even though the model itself is GA. The model does not support the Live API, image generation, or audio generation.
There is migration work. Google says to remove temperature, top_p, and top_k, replace thinking_budget with thinking_level, and stop using prefilled model turns. Applications using generateContent also need names and call IDs on function responses.
Gemini Managed Agents now use 3.8 Flash by default for the Antigravity agent and SDK. That makes this more than a new model option: existing managed-agent evaluations may need to be rerun because the underlying default changed.
My take: The stable endpoint is the headline, but the calendar is the decision. Store the introductory and January prices separately, meter thinking tokens, and pin an explicit model rather than inheriting a managed-agent default silently. The full Gemini 3.8 Flash guide has the complete pricing matrix and migration checklist.
2. Claude Fable 5.1 changes the economics of repeated agent context
Anthropic launched Claude Fable 5.1 on September 1 as a generally available model for long-running coding and knowledge work. It is available through the Claude API, consumer and team plans, and supported cloud marketplaces. The direct API model ID is claude-fable-5-1.
Base token prices remain premium at $10 per million input tokens and $50 per million output tokens. The material price change is prompt-cache reads: $0.25 per million tokens, down 75% from Fable 5. Anthropic estimates that change reduces typical token-billed workload cost by about 25% and highly agentic workload cost by as much as 45%.
Those are Anthropic estimates, not a universal discount. Savings depend on stable reusable prefixes, cache-hit rates, the ratio of cached input to fresh input, and how much output the agent produces. A workload that constantly changes its context or emits long answers will not realize the headline reduction.
The Fable 5.1 announcement also introduces an operational wrinkle. Fable 5.1 and the restricted Mythos 5.1 use the same underlying model but different safeguards. Fable queries flagged in advanced cybersecurity or biology can be routed to an Opus model. Anthropic says API customers must configure the new Fallback API for this behavior, and they are not charged Fable rates for rerouted requests.
Data retention still needs review. Fable requires 30-day retention for safety monitoring by default. Anthropic plans to phase in customer-controlled Enterprise Frontier Safeguards later this fall, with temporary zero-data-retention access for eligible customers until then.
My take: Fable 5.1 is not a new default for ordinary API calls. It is a selective escalation model for the hardest repository-scale and long-running work. The cache cut makes persistent agents more plausible, but only if you design prompts for reuse and measure real cache hits. See the Fable 5.1 developer guide for specs, access, and deployment boundaries.
3. VS Code separates portable agent sessions from portable agent extensions
VS Code 1.135 brings two related ideas into focus: a dedicated Agent Host for persistent sessions and Agent Plugins 1.0 for packaging reusable agent capabilities.
The distinction matters:
| Layer | What it standardizes |
|---|---|
| Agent Host Protocol (AHP) | Communication between an agent host and clients that display or control sessions |
| Agent Plugins 1.0 | Packaging for portable skills and MCP server configuration |
| MCP | Connections from agents to external tools and data |
Microsoft’s Agent Host architecture moves supported agent runtimes into a dedicated process. A session can continue independently of the editor window, multiple clients can stay synchronized, and the host can run beside a remote workspace. Copilot, Claude, Codex, and other harnesses can keep their own agent loops while adapters expose a common session model.
AHP is open and uses JSON-RPC, but it does not make every agent interchangeable. The protocol standardizes the client-facing session, not reasoning, context management, permissions, or tool behavior. Microsoft’s documentation also says the Agent Host and AHP remain under active development.
Agent Plugins 1.0 solves a different portability problem. The official plugin documentation defines a root plugin.json, skills under skills/, and portable MCP configuration in mcp.json. Client-specific commands, custom agents, rules, and hooks can live in reverse-domain namespaces that unsupported clients ignore.
That is a useful compatibility boundary. Skills and MCP servers can travel between compatible clients, while hooks and interface-specific behavior remain explicitly client-specific. VS Code continues to recognize Copilot, Claude, and legacy OpenPlugin layouts, so existing packages do not have to migrate immediately.
My take: Do not collapse AHP, Agent Plugins, and MCP into one “universal agent standard.” Together they cover session transport, extension packaging, and tool connectivity. Separately, each has a narrower contract developers can actually test. Our VS Code Agent Host and Agent Plugins 1.0 guide maps the formats and portability limits in detail.
4. Copilot content exclusions finally reach the app and CLI
GitHub made content exclusions generally available in the GitHub Copilot app and Copilot CLI on September 2. Enterprise, organization, and repository administrators can configure excluded paths that those agentic clients will not use as context.
Content exclusion was already generally available in IDE integrations. Extending it to the desktop app and terminal closes a practical governance gap: the same developer could previously move from an IDE surface with exclusions into a more autonomous client whose context controls differed.
The GitHub announcement says availability is limited to Copilot Business and Enterprise customers. Teams should verify effective rules at all three scopes rather than assuming an IDE policy automatically covers every client.
Exclusion is a context boundary, not a complete security boundary. An agent can still encounter secrets in command output, generated files, dependency logs, external tools, or data copied into a prompt. Repository permissions, sandboxing, secret scanning, and pull-request review remain necessary.
My take: This is not flashy, but it may be the most immediately actionable release of the week for enterprise teams. Inventory which Copilot surfaces developers use, test exclusions with canary files, and document what the rules do not cover. The broader GitHub Copilot app guide explains the app’s workspace and pull-request boundaries.
Quick hits
- IBM Granite 4.2 8B: IBM released an Apache-2.0 dense reasoning model with 128K context and selectable thinking modes. It is a practical self-hosting candidate, but IBM’s announcement does not by itself confirm a matching managed watsonx endpoint. See our Granite 4.2 8B guide.
- Mercury 2.5 Preview: Inception opened early access to a 260K-context diffusion language model with an OpenAI-compatible API. The preview status and provider-specific promotional pricing make it an evaluation target, not a frozen production dependency. See our Mercury 2.5 Preview guide.
- Copilot code review approvals: GitHub now lets Copilot code review approve pull requests when it finds no blocking issues. Treat an AI approval as one signal in branch protection, not a substitute for required human review on sensitive repositories.
That’s it for this week. Next Thursday: early production cost data for Gemini 3.8 Flash, whether Fable 5.1’s cache savings hold up in real agent loops, and which clients adopt the new plugin format.
Want this in your inbox? Subscribe to AI Dev Weekly.
Previous issue: AI Dev Weekly #23
FAQ
Is Gemini 3.8 Flash ready for production?
Yes. Google labels gemini-3.8-flash generally available with a stable model ID. Individual capabilities such as computer use remain in preview, and introductory pricing ends on December 31, 2026.
Did Claude Fable 5.1 get cheaper?
Base input and output prices are $10 and $50 per million tokens. Cache reads fell to $0.25 per million tokens, so workloads that repeatedly reuse large cached prefixes can become materially cheaper. Measure your own hit rate before assuming Anthropic’s estimated savings.
Are Agent Plugins 1.0 the same thing as MCP?
No. Agent Plugins package skills and MCP server configuration. MCP defines how an agent connects to tools and data. A plugin can contain MCP configuration, but the standards solve different problems.
Do Copilot content exclusions prevent all secret exposure?
No. They keep configured files from being used as Copilot context in supported clients. They do not replace repository permissions, sandboxing, secret scanning, output filtering, or code review.