Gemini 3.8 Flash: Google's Cheap Agentic Tier Tested (September 2026)
Google DeepMind released Gemini 3.8 Flash on September 2, 2026, and the headline is that it costs exactly what 3.7 Flash cost, 0.75 dollars per million input tokens and 3.75 per million output, while beating the older model on every benchmark Google published. It ships a 1 million token context, 64K output, multimodal input across text, image, audio, video, and PDF, and it is tuned for long horizon coding and autonomous agents. On DeepSWE v1.1 it reads 73.7 percent against 65.3 for 3.7 Flash, and on Terminal-Bench 2.1 it posts 89.4 percent. Google also claims it beats Claude Opus 5 on three of the benchmarks it reported. The catch worth naming up front is the price schedule: those rates hold through December 31, 2026, then double on January 1. This post tests whether the cheap tier is now good enough to be the default engine for real agent work, where it holds up, and where the flagship tiers still earn their keep.
Gemini 3.8 Flash: Google's Cheap Agentic Tier Tested (September 2026)
Google DeepMind released Gemini 3.8 Flash on September 2, 2026, and the pitch is unusually clean: same price as the model it replaces, better on everything Google measured. 3.7 Flash cost 0.75 dollars per million input tokens and 3.75 per million output; 3.8 Flash costs the same and beats it on every published benchmark. For anyone running agents at volume, a free capability bump on the cheap tier is more interesting than another flagship, because the cheap tier is where production economics are actually decided.
| Attribute | Gemini 3.8 Flash |
|---|---|
| Released | September 2, 2026 |
| Context window | 1,000,000 tokens |
| Max output | 64,000 tokens |
| Input | text, image, audio, video, PDF |
| Input price | 0.75 dollars per million tokens (through December 31, 2026) |
| Output price | 3.75 dollars per million tokens (through December 31, 2026) |
| Price after January 1, 2027 | doubles to 1.50 and 7.50 |
| DeepSWE v1.1 | 73.7 percent (3.7 Flash: 65.3) |
| Terminal-Bench 2.1 | 89.4 percent (3.7 Flash: 85.8) |
| OSWorld-2.0 | 59.0 percent (3.7 Flash: 50.6) |
The benchmark jump is real, and it lands where it matters
The improvements over 3.7 Flash are not cosmetic. DeepSWE v1.1 moves from 65.3 to 73.7 percent, an 8.4 point gain on an agentic software engineering benchmark. Terminal-Bench 2.1 goes from 85.8 to 89.4 percent. OSWorld-2.0, which measures a model driving a real computer environment, jumps from 50.6 to 59.0 percent, the largest single gain in the set. These are exactly the axes that matter for autonomous agents: can it run a multi step task in a shell or a desktop and finish it. An 8 point gain on the agentic benchmarks, at zero added cost, changes the calculus for teams that wrote off the Flash tier as too unreliable for real automation a generation ago.
Google also claims 3.8 Flash beats Claude Opus 5 on three of the benchmarks it published, and posts 54.9 percent on HLE-Verified against 53.6 for 3.7 Flash. Treat the cross vendor claim with the usual caution, because Google chose which three benchmarks to report, and a vendor reports the comparisons it wins. But a cheap tier model landing anywhere near a flagship on even a curated subset is a real shift. We tracked the previous generation of this same race in Gemini 3.7 Flash versus 3.6 Flash on the coding upgrade, and the pattern holds: the Flash tier is closing the gap to the flagships faster than the flagships are pulling away.
The pricing has a fuse on it
The number every procurement conversation needs to start with is the expiry. Gemini 3.8 Flash is 0.75 and 3.75 per million through December 31, 2026, and on January 1, 2027 both rates double to 1.50 and 7.50. Batch and Flex run at half those rates and Priority at 1.8x, so your effective number depends on the tier you route through, but the headline doubling is the planning risk. If you build a unit economics model on the launch pricing and ship a product whose margin depends on it, you have a step change coming in your cost base on New Year's Day that you did not underwrite.
The right response is not to avoid the model; it is to model both price regimes before you commit. Compute your cost per task at 0.75 and 3.75, then recompute at 1.50 and 7.50, and make sure the workload still clears margin at the higher rate. If it only works at the promotional price, you do not have a viable workload, you have a countdown. The doubling also narrows the gap to alternatives, so the competitive comparison you run today should be rerun against the January rates, because a model that is clearly cheapest now may be merely competitive in four months.
Where it holds up, and where it does not
Gemini 3.8 Flash is a strong default for high volume agentic work at moderate difficulty: ticket triage, tool calling loops, document extraction across its multimodal inputs, and the long horizon coding tasks it was explicitly tuned for. At 0.75 and 3.75 it is cheap enough to run speculatively, retry on failure, and still come out ahead of a flagship on cost per completed task. The 1 million token context means it can hold real repository or document context, and the multimodal input across audio, video, and PDF opens automation on document and media pipelines that a text only cheap tier cannot touch.
Where it does not hold up is the top of the difficulty curve. An 8 point gain over 3.7 Flash is meaningful, but 73.7 percent on DeepSWE v1.1 still means roughly one in four agentic tasks does not complete, and on the hardest reasoning and multi step problems the flagship tiers open a gap that matters when a failure is expensive. The production pattern that wins here is the same cascade we recommend across the cheap tier: route the bulk to 3.8 Flash and escalate only the failures to a flagship like GPT-5.6 Sol or Claude Opus 5. We laid out the cheap tier selection logic in detail in the cheap tier shootout across Haiku, Gemini Flash, and GPT mini, and the throughput and cost framing for the flagship escalation target in where GPT-5.6 Sol actually wins. The cheap tier is not a replacement for the flagship; it is the layer that keeps you from paying flagship rates on tasks that never needed them.
FAQ
How much does Gemini 3.8 Flash cost? 0.75 dollars per million input tokens and 3.75 per million output through December 31, 2026. On January 1, 2027 both rates double to 1.50 and 7.50. Batch and Flex are half those rates, Priority is 1.8x.
Is Gemini 3.8 Flash better than 3.7 Flash? Yes, at the same price. It beats 3.7 Flash on every benchmark Google published, including DeepSWE v1.1 (73.7 versus 65.3), Terminal-Bench 2.1 (89.4 versus 85.8), and OSWorld-2.0 (59.0 versus 50.6).
Can Gemini 3.8 Flash replace a flagship model for agents? For high volume, moderate difficulty work, often yes. For the hardest reasoning and multi step tasks it still trails the flagship tiers, so the reliable pattern is to run 3.8 Flash as the default and escalate only failures to a flagship.
What can it take as input? Text, image, audio, video, and PDF, with a 1 million token context window and 64K output, which makes it usable for multimodal document and media automation, not just text.
When this applies to your stack
A cheap tier that just got materially better at agentic work is the single highest leverage change most teams can make to their inference bill, and most teams miss it because they are watching the flagship launches. The work is to figure out how much of your volume 3.8 Flash can carry, where the escalation threshold to a flagship sits, and whether your margin survives the January price doubling. That is exactly the routing and cost engineering we do when we help teams design an agent stack that is cheap by construction rather than cheap by hope. If you want that modeled against your real traffic and both price regimes, our AI integration and agent engineering practice can run it.
Sources: Gemini 3.8 Flash pricing and benchmarks, OpenRouter, Gemini 3.8 Flash specs and pricing, llm-stats, Gemini 3.8 Flash for coding agents, Enterprise DNA.
More from the lab.
vLLM vs SGLang: Choosing a Production LLM Inference Engine in Late 2026
vLLM and SGLang are both open source, both built to serve LLMs at production scale, and both frequently recommended for the exact same job, which is exactly why teams keep picking the wrong one. The real difference is not a benchmark number. It is which attention and caching architecture matches your actual traffic pattern, and most teams never check that before they commit infrastructure budget to one or the other.
Ollama's MLX Backend Beats Metal by 93 Percent: Apple Silicon Local Inference Tested (October 2026)
Ollama's 0.19 release notes say it plainly: on Apple Silicon, Ollama is now built on top of Apple's MLX framework to take advantage of its unified memory architecture, rather than the Metal backed llama.cpp path it shipped for years. On an M4 Max running Qwen3.5, that swap took decode speed from 57.8 tokens per second to 111.4, a 93 percent improvement, and prefill from 1,154 to 1,810 tokens per second. The catch is in the fine print: the MLX runner is in preview, it needs 32GB or more of unified memory to activate at all, and as of this release it only accelerates one model family. Here is what actually changed, what it costs to turn on, and where it still falls back silently to the slower path.
DGX Spark vs Mac Studio M5 Ultra: Local LLM Inference Benchmarked (October 2026)
The M5 Ultra Mac Studio shipping in September 2026 is a quad die chip with 1.2 terabytes per second of memory bandwidth and up to 256GB of unified memory, with a 512GB configuration arriving in late October. Against a single NVIDIA DGX Spark, it is not close: on a 27B dense model, the Mac decodes at roughly four times the speed. Cluster two DGX Spark units together over NVIDIA's GX10 networking instead, and the comparison flips for a different workload entirely, concurrent multi user serving. Neither machine is the right answer for every workload, and the decision comes down to whether you are serving one user or several at once.