Claude Agents + Chat SDK, Gemini cuts tokens 17% — Dev Signal
Claude Agents + Chat SDK, Gemini cuts tokens 17%
Share:
Tool of the Week
Claude Managed Agents now work with Chat SDK
Server-side agent loop (model, tools, state, research) paired with Chat SDK's type-safe handler eliminates agent infrastructure work and ships to 30+ platforms via adapters.
Removes the need to build and host your own agent loop, session management, or platform integrations. Drop Anthropic credentials into Chat SDK handlers and deploy to Slack, WhatsApp, Discord, or web without webhooks or database.
Replaces DIY agent wiring and multi-platform adapter code. Requires Anthropic API key and Chat SDK setup. Start now: quickstart builds a working research analyst in the browser with zero platform registration.
“Claude Managed Agents handles the agent loop server-side, including the model, tools, session state, and sandboxed web research”
“Chat SDK gives that agent a chat interface through a single type-safe handler, with adapters that carry it to Slack, WhatsApp, and more”
“No database to run: The Managed Agents session stores the conversation, so the sidebar, transcript, and replay read from it, no server-side state of your own”
“Swapping a few lines in the handler moves your agent to Slack, Teams, Discord, WhatsApp, and 30+ other platforms”
Get issues like this in your inbox — free, every weekday.
Quick Signals
Gemini 3.6 Flash cuts token usage 17 percent
3.6 Flash reduces output tokens by 17% versus 3.5 Flash at lower cost per token ($7.50/1M), with 3.5 Flash-Lite hitting 350 tokens/second for high-throughput agentic workloads.
Lower token consumption and latency directly reduce agent inference costs and execution time in production systems. 3.5 Flash-Lite's throughput targets the scaling bottleneck for multi-agent workflows and document processing at scale.
3.6 Flash replaces 3.5 Flash for most agent workloads; no migration required beyond endpoint changes. 3.5 Flash-Lite replaces 3.1 Flash-Lite and 3 Flash for latency-critical paths. Ready to deploy now via Gemini API. Requires benchmarking against your agentic task mix to confirm token savings.
“reduces output token usage by 17% compared to 3.5 Flash”
“350 output tokens per second according to the Artificial Analysis Index”
“At $1.50/1M input tokens and $7.50/1M output tokens”
“3.6 Flash delivers higher precision with fewer unwanted code edits and reduced execution loops, as seen in DeepSWE (49% vs. 37%)”
“3.5 Flash-Lite is the fastest model in the 3.5 series”
OpenAI releases Codex Security CLI for vulnerability scanning
New @openai/codex-security package scans repos and CI pipelines for vulnerabilities via CLI or TypeScript SDK, requires Node.js 22+ and Python 3.10+.
Integrates security scanning directly into developer workflows and CI without switching tools. Supports both ChatGPT and API-key auth, with straightforward state management for reproducible scans across environments.
Open benchmark evaluates vulnerability detection across 25 models with scoring weighted 2:1 on recall over precision; cheapest competitive option is Grok 4.5 high ($5.60) vs GPT-5.6 Sol xhigh ($55.98) with top score 35.58.
Developers can now compare vulnerability detection cost-to-performance across models to optimize their automated security scanning pipelines, replacing expensive frontier-model-only approaches with cheaper multi-model strategies that fit their codebase complexity and budget constraints.
Replaces manual model selection guesswork with empirical recall/precision/cost data. Requires integrating deepsec tool and selecting model tier by cost-benefit tradeoff. Ready now—benchmark is live and public. Caveat: best model only finds 30.7% of vulnerabilities; no single pass is comprehensive.
“the benchmark is run three times and the data published in the report is the median of the three runs”
“we don't disclose the repository, the commit, the files, or the findings, so there is nothing for models to train against”
“higher price does not buy proportionally more”
“Kimi K3, from Moonshot AI, ranks eighth at a score of 17.56 for $12.38 on the high setting, half the top score for about a fifth of the cost”
“Anthropic's most capable model, Fable 5, is absent because it declines security work, including defensive tasks”
“weighting recall (R) twice as much as precision (P), because missed vulnerabilities will go unfixed, while false positives don't make your codebase less secure”
Replaces ad-hoc security audits or separate SAST tools. Requires Node 22+, Python 3.10+, and Codex Security access. Ready to try now if you have OpenAI API access—install via npm, authenticate, run `scan .` in seconds. CI setup is standard env-var driven.
“Requires Node.js 22 or later, Python 3.10 or later, and access to Codex Security”
“Scan repositories, review changes, track findings over time, and run security checks in CI”
“If both a ChatGPT sign-in and an API key are available, interactive scans ask which credential to use”
AI Gateway adds WebSocket support for Responses API
Replace HTTP round-trips with persistent WebSocket connections; send only new input plus previous_response_id to cut context re-transmission overhead.
Multi-turn agentic workflows with 20+ tool calls see ~40% faster end-to-end execution. Reduces bandwidth and latency for stateful AI interactions, critical for real-time agent loops.
Replaces sequential HTTP calls with stateful WebSocket frames. Requires routing to wss://ai-gateway.vercel.sh/v1/responses and Bearer auth. Ready now—live on AI Gateway with store=false and ZDR compatible. Worth adopting if you're building multi-turn agents with frequent tool calls.
“up to ~40% faster end-to-end execution on WebSockets for agentic rollouts with 20 or more tool calls”
“keep a persistent connection open and continue each turn by sending only new input items plus `previous_response_id`”
“compatible with `store=false` and Zero Data Retention”
websocketopenai-apiai-gatewayagentslatency
LFM2.5-Encoders match larger models at 8K tokens
Two new bidirectional encoders (230M, 350M params) run document-scale classification and routing on CPU at 3.7× faster throughput than ModernBERT, scaling efficiently to 8,192-token context.
Replaces the CPU inference bottleneck for high-volume NLP tasks—intent routing, PII detection, policy checking—letting you deploy classifiers on existing hardware without GPU. Backward-compatible via standard transformers API.
Drop-in replacement for BERT-class encoders on CPU-bound workloads. Requires fine-tuning for downstream tasks (included tutorial targets legal documents). Ready now: open-weight models on Hugging Face, live demos run in browser, pip install transformers covers setup. Worth adopting if you own classification/routing latency on CPU.
“match or beat larger encoders on GLUE, SuperGLUE, and multilingual tasks”
“8,192-token context with latency that grows slowly as inputs get longer”
“about 3.7× faster than ModernBERT-base at long context”
“At 8,192 tokens, ModernBERT-base takes over a minute and a half per forward pass versus about 28s for LFM2.5-Encoder-230M”
“you can scan or classify a full contract, transcript, or long support thread in under 30 seconds on a laptop CPU”
Vercel adds WebSocket support for Python functions
Python Functions on Vercel now support ASGI/WSGI WebSockets natively, letting you stream responses and handle bidirectional client-server communication without external infrastructure.
Eliminates the need for separate WebSocket servers when building real-time AI features (streaming chat, live updates) on Vercel. Simplifies deployment for FastAPI, Django, and Flask apps that require persistent connections.
Replaces the need to self-host or use third-party WebSocket providers for Python on Vercel. Requires Python 3.9+ and an ASGI or WSGI framework. Ready to use now—starter examples provided for FastAPI and Flask. Worth adopting if you're already on Vercel and need real-time features.
“WebSockets enable bidirectional communication between client- and server-side code”
“Both ASGI and WSGI applications are supported, including frameworks like FastAPI, Django, and Flask”