Three things builders should know
- The GA is real and dated.
gemini-3.6-flashandgemini-3.5-flash-liteboth reached stable status on 21 July 2026, per Google's own Gemini API changelog. If your team spent July heads-down shipping, this is the release to catch up on. - Output got cheaper where it matters. 3.6 Flash lands at $1.50/M input and $7.50/M output — the same input price as 3.5 Flash but $1.50/M less on output. Google also says the model consumes 17% fewer output tokens than 3.5 Flash, which compounds the saving on verbose agentic loops.
- Flash-Lite is the actual subagent play. At $0.30/M input and $2.50/M output, Gemini 3.5 Flash-Lite is positioned explicitly for high-throughput, low-latency tasks such as agentic search and document processing — the jobs you fan out to workers, not the ones you give your orchestrator.
What actually shipped on 21 July
Google's changelog entry is short but dense. The stable gemini-3.6-flash "features improved token efficiency and code/agentic planning capabilities", while gemini-3.5-flash-lite "offers a low-latency, highly cost-effective subagent option" — Google's wording, not ours, and a fairly direct admission of where these models are aimed. Both are live in the Gemini API via Google AI Studio and Android Studio, in the Gemini Enterprise Agent Platform for enterprise accounts, and in the consumer Gemini app.
The launch coverage from 9to5Google adds Google's own benchmark comparisons against 3.5 Flash: DeepSWE 49% versus 37%, MLE Bench 63.9% versus 49.7%, GDPval-AA 1421 versus 1349, and OSWorld-Verified 83% versus 78.4%. Flash-Lite's headline number is Terminal-Bench 2.1 at 54% versus 31% for its predecessor. The knowledge cutoff moves forward to March 2026. Treat all of these as vendor-reported until independent harnesses publish their runs — but the direction of travel, stronger agentic and coding behaviour at the cheap end, is consistent across every figure Google published.
Two quieter changes deserve your attention. First, the changelog deprecates the sampling parameters temperature, top_p and top_k — if your prompt pipeline sets these on every request, start planning their removal now rather than when they hard-fail. Second, gemini-robotics-er-1.6-preview shuts down on 31 August 2026, replaced by the new Gemini Robotics ER 2 preview endpoints announced on 30 July. Robotics teams in Bengaluru and Bristol alike have three weeks to migrate.
The benchmark deltas above are Google's published comparisons, not third-party results. Before you re-route production traffic, run 3.6 Flash against your own eval set. Vendor benchmarks tell you a model improved; they do not tell you it improved on your tasks.
The pricing maths: cheaper exactly where subagents spend
Subagent workloads are output-heavy. A worker that reads a brief, searches, extracts and reports back typically emits far more tokens than the orchestrator that briefed it. That makes the output-price cut the interesting number here — and the claimed 17% token-efficiency gain a multiplier on top of it, if it holds on your traces.
| Model | Input $/M | Output $/M | Cache read $/M | Batch (in/out) |
|---|---|---|---|---|
| Gemini 3.6 Flash | $1.50 | $7.50 | $0.15 | $0.75 / $3.75 |
| Gemini 3.5 Flash | $1.50 | $9.00 | $0.15 | $0.75 / $4.50 |
| Gemini 3.5 Flash-Lite | $0.30 | $2.50 | $0.03 | $0.15 / $1.25 |
| Claude Haiku 4.5 | $1.00 | $5.00 | — | — |
Gemini figures are from Google's Gemini API pricing page (context-cache storage is an additional $1.00 per million tokens per hour); Claude Haiku 4.5 is from Anthropic's pricing page. On paper Haiku 4.5 undercuts 3.6 Flash on both sides of the ledger, and OpenAI's mini-class models — GPT-5.4 mini is commonly listed around $0.75/M input and $4.50/M output on third-party pricing trackers, though you should confirm against OpenAI's own page — sit in a similar band. Sticker price is not the decision, though. As we argued when 3.5 Flash's cost-per-task economics first shook out, what matters is the cost of a completed task: a cheaper model that needs three retries or a supervisor pass loses to a dearer one that finishes first time. Our guide to LLM unit economics walks through how to measure that properly.
Subagent fleets share a lot of prompt: the same system prompt, tool schemas and task template go to every worker. At $0.15/M cache reads on 3.6 Flash — a tenth of fresh input — caching that shared prefix is the difference between a fan-out of 20 workers costing 20× and costing roughly 3–4×. Warm the cache once, then spawn.
Where a Flash tier fits in an orchestrator stack
The now-standard production pattern puts a frontier model at the top of the stack — planning, decomposing, judging — and pushes the reading-heavy, parallelisable work down to cheap workers. If you run that pattern with Claude Code subagents or any framework equivalent, the GA gives you two distinct slots to fill:
- 3.6 Flash as the mid-tier worker. Multi-step subagent tasks that involve tool calls, code, or judgement: extract-and-verify pipelines, agentic search with follow-ups, test-writing workers. This is where the improved agentic planning claim, if it survives your evals, earns the $7.50 output rate.
- 3.5 Flash-Lite as the swarm layer. Single-purpose, high-volume calls: classification, routing, per-document extraction, first-pass triage. At $0.30/$2.50 you can afford to be wasteful with parallelism and strict with verification — send three workers, take the majority vote, and still spend less than one mid-tier call.
The discipline that makes either slot pay off is routing, not model choice. A routing-and-cascade setup that sends only genuinely hard tasks up the stack routinely cuts spend by more than any single model swap. And if you have never audited where your multi-agent tokens actually go, do that first — the breakdown is usually not where teams expect, with orchestrator re-reads and duplicated context dwarfing the worker calls everyone worries about.
"We route on task shape, not model loyalty. Anything that is one read and one structured answer goes to the cheapest tier we trust that week; anything with a loop in it goes one tier up. When a new Flash or Haiku ships, we re-run the same 200-trace eval and the router config changes, not the code."
— Arun, Verified Builder · London, UKEvery article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →The honest comparison: Haiku-class and mini-class rivals
Google is not entering an empty room. Anthropic's Haiku 4.5 has been the default "fast worker" in many Claude-centric stacks precisely because it shares tooling, prompt conventions and caching with the frontier models above it. OpenAI's mini tier holds the same position in GPT-based stacks. And the open-weight flank is crowded: DeepSeek's V4 Flash went GA at the end of July with agent benchmarks aimed squarely at this segment, at prices the closed labs will not match.
So the practical question is not "which is best?" but "what does switching cost you?" Three qualitative points to weigh, since head-to-head independent numbers are still thin:
- Stack coherence has real value. A same-family worker inherits your caching, tool-schema quirks and failure modes. Mixing vendors mid-stack means two sets of both. The saving has to clear that bar.
- Token efficiency changes the comparison. If Google's 17% output-reduction claim holds on your workload, 3.6 Flash's effective output cost approaches $6.20/M against Haiku's $5.00 — close enough that quality on your tasks, not price, decides it.
- Context and surface differences matter at the edges. Haiku 4.5 carries a 200K context window; Google has not published a changed window for 3.6 Flash in the changelog, so verify limits against your longest worker payloads before committing.
What Indian and UK teams should test this week
A concrete checklist, in the order we would run it:
- Re-run your subagent eval set on
gemini-3.6-flashandgemini-3.5-flash-lite, logging cost per completed task alongside pass rate. Twenty representative traces beat any benchmark table. - Trial the swarm pattern on Flash-Lite for one high-volume step — invoice-field extraction for a Mumbai fintech, KYC document triage for a London one. Three cheap parallel calls plus a vote is often more reliable than one mid-tier call.
- Move batch-tolerant jobs to batch pricing. Overnight enrichment, backfills and re-embeds do not need interactive rates; $3.75/M output on 3.6 Flash batch is half the live price.
- Audit for the deprecated sampling parameters. Grep your codebase for
temperature,top_pandtop_kon Gemini calls and plan their removal. - If you touch robotics, migrate off
gemini-robotics-er-1.6-previewbefore the 31 August shutdown.
None of that needs a big-bang migration. The cheapest way to run the comparison is shadow traffic: leave your existing worker tier serving production, mirror a slice of live requests to gemini-3.6-flash, and diff the two sets of outputs offline against the grader you already trust. You get your real task distribution rather than a curated eval set, no user-facing risk if the new model behaves oddly on an edge case, and a cost figure that includes the retries — which is the only number that ever settles a routing argument. A week of mirrored traffic is usually enough to see whether the token-efficiency claim survives contact with your prompts.
For teams on either side of the corridor, the procurement angle differs. Indian startups paying in rupees feel the output-price cut directly on thin-margin consumer products, where a few paise per request decides whether a feature ships. UK teams, more often bound by enterprise data-residency and vendor-review processes, will care that the same stable model IDs now flow through the Gemini Enterprise Agent Platform — a GA label is what unblocks the compliance sign-off that a preview endpoint never gets.
One more signal worth filing away: in the launch-day coverage cited above, Google told 9to5Google it has "already started our most ambitious pre-training run yet, for Gemini 4". The cheap tier you standardise on today is also a bet on which frontier family sits above it next quarter. Choose the worker tier you can swap out — and the harness that makes swapping cheap.