Gemini 3.6 Flash and 3.5 Flash-Lite available on AI Gateway
Gemini 3.6 Flash reduces token consumption and model calls for coding/web tasks; 3.5 Flash-Lite handles scoped agentic subtasks—both callable via unified AI SDK with cost tracking and failover built in.
Token efficiency directly cuts inference costs and latency in production agents. Unified routing through AI Gateway consolidates usage tracking, budgets, and failover across providers without platform markup, eliminating need for custom middleware.
Replaces ad-hoc Google API calls with standardized AI SDK routing. Requires setting `model` parameter to `google/gemini-3.6-flash` or `google/gemini-3.5-flash-lite`. Worth trying now if you're already on Vercel or using AI SDK—zero friction adoption, no pricing changes for existing users.
“Gemini 3.6 Flash improves quality across coding, agentic tasks, and web development while consuming fewer tokens and making fewer model calls”
“AI Gateway provides a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations”
“AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference”
Get issues like this in your inbox — free, every weekday.
Quick Signals
Gemini 3.6 Flash cuts tokens by seventeen percent
3.6 Flash reduces output token usage 17% versus 3.5 Flash at lower cost per token; 3.5 Flash-Lite hits 350 tokens/sec for high-throughput agentic loops.
Token efficiency directly impacts agent inference cost and latency at scale. Fewer output tokens per task means cheaper multi-step workflows and faster real-time agent execution—critical for production deployments where cumulative overhead compounds.
3.6 Flash replaces 3.5 Flash for production agents; drop-in API swap with pricing $1.50/1M input, $7.50/1M output. 3.5 Flash-Lite ($0.3/$2.5) trades quality for throughput on high-volume tasks. Both ship now via Gemini API. Worth migrating existing agents immediately if cost or latency is a bottleneck.
“reduces output token usage by 17% compared to 3.5 Flash”
“350 output tokens per second according to the Artificial Analysis Index”
“$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%)”
3.6 Flash reduces output tokens while improving coding performance and lowering cost to $7.50/1M, making agentic workloads cheaper to run at scale.
Token efficiency directly reduces inference costs for production agents. Lower latency (3.5 Flash-Lite at 350 output tokens/s) and improved reasoning steps mean faster iteration on multi-step workflows without code bloat.
Data Point
ProGraph matches FullContext on multi-hop memory
Two-layer memory architecture using profile expansion and compression residuals achieves 80.1% on MemHop benchmark without explicit knowledge-graph construction.
Current LLM agent memory systems fail at multi-hop reasoning across sessions. ProGraph trades graph indexing overhead for substring traversal of LLM-generated profiles, reducing API calls while handling 5-hop queries—directly improving agent reliability in long-running applications.
Replaces explicit knowledge-graph memory systems (HippoRAG, A-Mem, Mem0). Requires implementing profile expansion logic and compression residual extraction alongside standard context storage. Worth prototyping now if building agents with multi-session memory; code and benchmark released.
“profile expansion -- substring-matched traversal of entity names that naturally appear in LLM-written profile narratives, a minimal alternative to explicit knowledge-graph construction”
“compression residuals -- exact dates, quantities, and named items co-extracted with each profile update at zero extra API cost”
“ProGraph averages 80.1% on MemHop (matching the FullContext reference) and 78.4% on LoCoMo (exceeding FullContext by 11.3pp)”
“Long-term memory is essential for LLM agents that interact across sessions, yet current memory benchmarks primarily evaluate single-hop recall, leaving multi-hop association largely unmeasured”
3.6 Flash replaces 3.5 Flash for most agentic tasks; 3.5 Flash-Lite replaces 3.5 Flash for high-throughput document processing and search. Pricing is lower, benchmarks show measurable gains. Try now if you're already on Gemini—no API changes required, just update model selection.
“reduces output token usage by 17% compared to 3.5 Flash”
“delivering 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 reduces the overall cost per agentic task”
“3.6 Flash delivers higher precision with fewer unwanted code edits and reduced execution loops, as seen in DeepSWE (49% vs. 37%)”
“Priced at $0.3/1M input tokens and $2.5/1M output tokens”
Open-weight MoE model with 1M context window and thinking mode lands on Vercel's unified API gateway for coding tasks, priced at provider rates with no platform markup.
Replaces multi-provider integrations with single API surface for agentic workflows; 1M context enables long-running test debugging and MLOps tasks without context switching. Benchmarks (78.5% SWE-bench Multilingual) suggest viable drop-in for code-focused agents.
Replaces direct Poolside API calls or competing long-context models (Claude, GPT-4). Requires one SDK line change to `poolside/laguna-s-2.1` or free tier variant. Ready now—but validate benchmarks against your specific task (SWE-bench skews toward repo-scale issues, not all coding patterns).
“supports a context window of up to 1M tokens and runs in thinking and no-thinking modes”
“70.2% on Terminal-Bench 2.1, 78.5% on SWE-bench Multilingual, and 59.4% on SWE-Bench Pro”
“specializes in agentic coding and long-running tasks, including writing and debugging code, running tests, building browser-based tooling, and working on MLOps pipelines and AI research”
“AI Gateway reflects provider pricing with no markup and does not charge a platform fee on inference”
AI Gateway adds service tier routing for latency-cost tradeoffs
Set `serviceTier` in `providerOptions.gateway` to route requests between `default`, `priority` (~1.8-2x cost), or `flex` (~0.5x cost) tiers; billing auto-adjusts based on actual tier used.
Developers can now optimize request handling without code restructuring—interactive endpoints get priority queuing while background jobs use cheaper capacity. Tier configuration is unified across OpenAI, Gemini, and other providers, reducing boilerplate.
This replaces manual provider-specific configuration for throughput/latency tuning. Requires single line addition to `providerOptions` in AI SDK or compatible APIs. Production-ready now; best-effort fallback to default tier on capacity constraints means no failure risk. Worth testing immediately if you have mixed latency requirements.
“Service tiers let you optimize for latency, throughput, and cost per request to match your use case”
“Pick a faster tier for interactive workloads (less queueing, higher token throughput), or a lower cost tier for background jobs that can tolerate more latency”
“Set `serviceTier` under `providerOptions.gateway`”
“Cost relative to default: ~1.8-2x default pricing for priority, ~0.5x default pricing for flex”
“Service tier is serviced on a best-effort basis: if a tier can't be applied, the request runs on the default tier at the default rate”
Vercel now bundles precompiled .pyc files with Python functions, eliminating runtime compilation overhead and cutting cold starts by ~53% for median workloads.
Cold start latency directly impacts user experience for event-driven Python workloads. Automatic bytecode precompilation removes a structural bottleneck without requiring code changes, making baseline performance faster for all Python deployments on the platform.
Replaces the runtime compilation step that Python normally performs on first import. Requires nothing—Vercel handles it automatically at build time. Worth adopting now if you deploy Python functions on Vercel; negligible downside for functions already near bundle limits.
“cold starts for the median-sized function dropped from 2.8s to 1.3s”
“When Python imports a module without cached bytecode, it parses and compiles the source before executing it”
“Vercel automatically adds as much precompiled bytecode as fits in the function bundle”