AI Dev Weekly #26: Agents API, Full-Duplex Voice, SWE-2 and Cost-Aware Copilot
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 model is no longer the whole product. OpenAI now sells the managed agent loop behind Codex and a separate voice layer that can keep talking while an agent works. Cognition’s SWE-2 focuses on completing coding tasks with fewer detours. Copilot lets developers express a cost-versus-quality preference instead of choosing every model manually. The common question: which parts of the agent stack should you own, and which should you delegate?
This issue includes the September 10 releases that landed after last Thursday’s edition, alongside updates through September 17.
1. OpenAI puts the Codex harness behind the Agents API
OpenAI released the Agents API in public beta on September 10. It is a managed runtime for durable agent work, not another name for the Agents SDK. OpenAI handles session orchestration, context compaction, and recovery; your application supplies tools and controls how the agent fits into your product.
The official overview separates four concepts: the agent configuration, its execution environment, a durable session, and the events and items exchanged during that session. Progress can arrive through streaming or webhooks, and a session can receive follow-up work rather than restarting from an empty context.
The architecture guide makes the execution boundary particularly clear:
- No environment: the harness can use remote MCP connections and application-handled function tools, without a built-in shell or workspace files.
- OpenAI-hosted environment: OpenAI provisions the sandbox for code execution, file edits, and artifacts.
- Self-hosted environment: your infrastructure connects an executor, while OpenAI still runs the harness. You own provisioning, reconnection, shutdown, and file preservation.
That last option is not self-hosting the complete agent. Moving execution into your VPC does not move the managed orchestration layer there.
Pricing combines the selected model’s API rates, standard tool charges, and standard container rates for OpenAI-hosted sandboxes. Budget the whole run, including subagents and retries, rather than treating a session as one model response.
The security documentation also warns that agent-generated code can read credentials exposed to its environment. Keep the application API key outside the sandbox, restrict outbound destinations, and broker third-party credentials where possible. A sandbox does not make a secret safe after you inject it into the process.
My take: This is a meaningful build-versus-buy shift. Session recovery and compaction are expensive plumbing to maintain, so renting them can make sense. Keep authoritative business state and irreversible-action approvals in your own application, however. Start with a read-only investigator before a remediation agent. Our Cursor Projects vs OpenAI Agents API comparison explains the difference between buying a coding workspace and embedding an agent runtime in your product.
2. GPT-Live-1 separates the conversation from the work
GPT-Live-1 is generally available in the API, also from September 10. Its main architectural change is full duplex: the voice model can listen while speaking, and the spoken conversation can continue while a backend agent handles reasoning and tools.
OpenAI’s getting-started guide contrasts this with Realtime’s single model for speech, reasoning, and tool selection. With Live, you choose the backend independently. A user can ask about an order, correct a detail while a lookup runs, and receive the result without waiting through a silent tool round trip.
The delegation guide offers two modes:
- Responses delegation: Live prepares requests for a configured OpenAI backend and returns its results to the conversation.
- Client delegation: your application assembles context, runs its own agent or service, and returns the result. This is the path when you need custom routing or validation before backend results reach the voice model.
Client delegation has a subtle integration requirement: delegation events contain metadata, not the task text. Build backend requests from transcript events and application state. Do not assume the event is a ready-made command payload.
The model page lists voice sessions at $0.05 per minute, billed per second. Backend model and tool usage is additional. A ten-minute voice session therefore costs $0.50 for the voice layer before reasoning, tools, or your infrastructure.
Permissions, confirmation, and task state remain application responsibilities in both modes. Reviewing backend results also does not guarantee the voice model stays silent while validation runs. Test what the assistant says before an action is confirmed, not just whether the tool ultimately succeeds.
My take: Voice becomes an interface over an agent rather than a tightly coupled speech pipeline. That is especially useful for support workflows and hands-free developer collaboration. My first evaluation would test interruptions, changing instructions during a slow lookup, and whether the assistant accurately distinguishes “requested” from “completed.” Natural conversation is valuable only if the task state stays truthful. This is a different problem from the speech-to-text decision in our Gemini Transcribe vs Whisper guide.
3. Cognition SWE-2 targets fewer turns, not just a higher score
Cognition introduced SWE-2 on September 10, launching in Devin Desktop and CLI with rollout to Devin Web and Fusion. The launch post says the model is post-trained from Kimi K3 and trains multiple reasoning-effort levels in one reinforcement-learning run.
The most useful vendor result is operational: SWE-2 at medium effort used 58% fewer turns and cost 81% less than SWE-1.7 on FrontierCode 1.1 Main. Cognition reports a 50.0% score on that benchmark for SWE-2, but only 27.3% on Terminal-Bench 4. The latter trails the strongest models in its comparison substantially.
These are publisher-reported evaluations with a mixture of public results and internal runs in model-specific harnesses. They are not an independent, uniform leaderboard. Cognition has not published standalone SWE-2 API pricing or downloadable weights in the announcement.
My take: The interesting optimization is knowing when to stop exploring and start editing. For a coding agent, fewer repository reads and redundant tool calls can matter more than marginal token-price differences. But “fewer turns” is not automatically “better”: it can also mean skipped verification. Evaluate a bounded bug fix, a migration, and a behavior-preserving refactor, recording accepted changes, missed tests, review corrections, and actual product usage. Our SWE-2 guide covers the access limits and benchmark caveats. Test the Devin capability you can actually use, not an imagined standalone model endpoint.
4. Copilot Auto gains a cost-versus-quality control
GitHub added three Auto routing preferences on September 14: Efficiency for lower cost, Balance for cost, quality, and latency together, and Intelligence for quality. The release announcement says the feature is rolling out in VS Code, Copilot CLI, and the Copilot app.
These are not three separate model catalogs. Each preference uses the same available pool, and routing evaluates the prompt. Intelligence can still select a small model for a simple docstring task. Usage is billed for the model actually selected, and paid subscribers retain a 10% Auto discount.
The Auto documentation adds two useful constraints. The router still respects plan access and administrator policies, and routing uses natural cache boundaries to avoid extra cache-related cost. Auto does not override an organization’s restricted-model policy.
A preference is also not a spending cap. Efficiency can choose a more capable model when the task needs one. Intelligence does not guarantee the largest model on every response. Inspect the selected model and actual usage before drawing conclusions from the label.
My take: This is a better abstraction than forcing every developer to maintain a mental model leaderboard. It also makes comparisons harder unless you record routing decisions. Run the same representative tasks under each preference and compare cost per accepted change, latency, and reviewer effort. Keep explicit model selection for evaluations that require reproducibility. Our updated Copilot app guide and usage-based billing explainer separate routing preference from billing policy.
Quick hits
- VS Code agent automations: Version 1.137 adds recurring hourly, daily, and weekly tasks in public preview. Start with maintenance reports or draft pull requests, and define what happens when runs overlap. The same weekly release notes put HydraFusion’s local/cloud/compound routing in the CLI’s experimental menu. Neither should be confused with the standard Auto preferences above.
- Copilot code review: It now resolves its own addressed comments during rereview and gains shell tools for builds, tests, and targeted scripts behind the agent firewall. GitHub’s update also introduces an agent ensemble for Lite reviews. More validation is useful; required human review remains a separate repository decision.
- API key governance: OpenAI added organization and project controls on September 15 to limit new key creation to service accounts, user-owned project keys, or neither. Organization restrictions take precedence and existing keys are unaffected. The API changelog also documents key expiry controls added September 10. Creation policy is not retroactive credential rotation.
That’s it for this week. Next Thursday: how managed agent sessions recover from real failures, whether full-duplex voice keeps task state honest, and what routing preferences cost on actual repository work.
Want this in your inbox? Subscribe to AI Dev Weekly.
Previous issue: AI Dev Weekly #25
FAQ
Is OpenAI’s Agents API the same as the Agents SDK?
No. The API provides an OpenAI-managed harness and durable sessions. The SDK provides primitives for application-owned orchestration. The new API remains in public beta.
Does GPT-Live-1’s price include the backend agent?
No. The voice layer costs $0.05 per minute, billed per second. Backend model and tool usage is charged separately.
Can I download Cognition SWE-2 or call a standalone API?
Neither downloadable weights nor standalone API pricing were published in the launch announcement. Evaluate SWE-2 through the documented Devin surfaces.
Does Copilot Efficiency guarantee a fixed bill?
No. It is a routing preference, not a cap. Billing follows the selected model, so measure actual usage and enforce budgets separately.