Gemini 3.8 Flash vs Claude Opus 5: Coding, Agentic Benchmarks, and the Token Cost Catch (2026)
Google shipped Gemini 3.8 Flash on September 2, 2026, at 0.75 dollars per million input tokens and 3.75 per million output, and it beats Claude Opus 5 on three of the benchmarks Google chose to publish. Opus 5 sits at 5 and 25 dollars per million and holds the top score on SWE-bench Verified. On sticker price the Flash tier looks like an easy win, but 3.8 Flash is built to burn more thinking tokens than the model it replaces, which means the real bill per task can climb well above what the headline rate implies. This post puts the two side by side on the numbers that compare, and works through when the cheaper model is actually cheaper.
Gemini 3.8 Flash vs Claude Opus 5: Coding, Agentic Benchmarks, and the Token Cost Catch (2026)
Google shipped Gemini 3.8 Flash on September 2, 2026, and the headline is a cheap model that punches up: it beats Claude Opus 5 on three of the benchmarks Google chose to publish, at 0.75 dollars per million input tokens and 3.75 per million output. Claude Opus 5, released July 24, sits at 5 and 25 dollars per million and holds a 96.0 percent score on SWE-bench Verified, one of the highest coding results any vendor has posted. The reflex is to read the price gap and pick Flash. The reflex is wrong often enough that it is worth doing the arithmetic, because 3.8 Flash is deliberately built to spend more thinking tokens than the model it replaces.
| Attribute | Gemini 3.8 Flash | Claude Opus 5 |
|---|---|---|
| Vendor | Anthropic | |
| Released | September 2, 2026 | July 24, 2026 |
| Input price | 0.75 dollars per million tokens | 5 dollars per million tokens |
| Output price | 3.75 dollars per million tokens | 25 dollars per million tokens |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| SWE-bench Verified | not the headline benchmark | 96.0 percent |
| DeepSWE v1.1 | 73.7 percent | Google claims Flash leads on 3 of its benchmarks |
| Terminal-bench 2.1 | 89.4 percent | (version drift, see below) |
What Gemini 3.8 Flash actually improved
Gemini 3.8 Flash is not a new base model. It is built on 3.7 Flash and made to work harder, spending more reasoning tokens per request to lift its scores. On the benchmarks Google published it beats 3.7 Flash across the board: DeepSWE v1.1 climbs to 73.7 percent from 65.3, Terminal-bench 2.1 to 89.4 from 85.8, and OSWorld-2.0 to 59.0 from 50.6. Those are real gains on agentic and computer use tasks, the kind of work where a cheap model that can actually finish a multi step job is worth a lot.
The pricing is aggressive on purpose. At 0.75 and 3.75 dollars per million tokens, Flash undercuts every frontier model by roughly an order of magnitude on the sticker. That rate holds through December 31, 2026, then both numbers double to 1.50 and 7.50 on January 1, 2027, so any cost model you build on the launch price has a shelf life. Batch and Flex rates run at half, and Priority at 1.8 times, so the effective rate depends on which serving tier you commit to.
Where Opus 5 still wins
Claude Opus 5 is the premium end of the market and it earns the position on the hardest coding work. A 96.0 percent SWE-bench Verified score is the number to beat for repository level bug fixing, and Anthropic holds it while keeping Opus 5 at the same 5 and 25 dollar rate as Opus 4.8. On tasks where a wrong answer is expensive, a review agent gating a merge, a migration script touching production data, a refactor across a large codebase, the accuracy gap pays for the price gap. We walked through the reasoning side of this matchup in Gemini 3.7 Flash versus Claude Opus 5 on hard math reasoning, and the pattern holds: Flash closes distance on the mid tier, Opus 5 pulls away at the frontier.
The benchmark version caveat matters here. Google reports Terminal-bench 2.1; other vendors report different versions of the same benchmark with different task sets and scoring. Reading 89.4 on one version as strictly better than a number from another version is a category error, and it is the most common way these comparisons go wrong in 2026. Compare within a version or do not compare at all.
The token cost catch
Here is the part the sticker price hides. Gemini 3.8 Flash lifts its scores by spending more thinking tokens, and Google itself recommends staying on 3.7 Flash for efficiency first workloads. Thinking tokens are output tokens, billed at the output rate. If 3.8 Flash burns two or three times the reasoning budget of 3.7 Flash to reach its higher scores, the effective cost per completed task rises by the same multiple, and the 3.75 dollar output rate stops being the number that describes your bill.
Work the example. Suppose a task takes 2,000 output tokens on 3.7 Flash and 5,000 on 3.8 Flash because the newer model reasons harder. At 3.75 dollars per million output tokens, that is 0.0075 dollars versus 0.01875 dollars per task, a 2.5 times increase that no line on the pricing page shows you. Against Opus 5 at 25 dollars per million, a task that lands in 1,500 output tokens costs 0.0375 dollars. So the real comparison is 0.01875 against 0.0375, a two times gap, not the ten times gap the input rates imply. If Opus 5 solves the task on the first try and 3.8 Flash needs a retry, the gap can close to nothing or invert. We take the full cost per task methodology apart in AI coding agent pricing for 2026, and the lesson is the same every time: model the cost on tokens actually spent to finish, not on the headline rate.
How to choose between them
Route by task, not by default. Gemini 3.8 Flash is the right call for high volume agentic work where individual errors are cheap to catch and retry: bulk code generation behind review, computer use automation, tool calling loops, anything where throughput and a low floor price dominate. Its DeepSWE and OSWorld gains mean it now finishes multi step jobs that 3.7 Flash would stall on, which is often worth more than a marginal accuracy point. Just measure the token spend under your real prompts before you commit, because the thinking token behavior can erase the price advantage on reasoning heavy requests.
Claude Opus 5 is the right call where correctness gates something expensive and the token count per task is bounded: the final review, the schema migration, the security sensitive change, the frontier coding task where 96.0 percent SWE-bench Verified is the difference between shipping and reverting. The clean architecture is a cheap default model with an expensive escalation path, and 3.8 Flash plus Opus 5 is a reasonable instance of exactly that pattern. We compared the cheap Flash tier on its own terms in Gemini 3.8 Flash as a cheap agentic tier, tested.
FAQ
Is Gemini 3.8 Flash better than Claude Opus 5? On the three benchmarks Google published, Google claims Flash leads. On SWE-bench Verified, Opus 5 holds a 96.0 percent score that Flash does not target. They are built for different jobs: Flash for cheap high volume agentic work, Opus 5 for hard coding where accuracy gates cost.
How much cheaper is Gemini 3.8 Flash? On sticker price, roughly seven times cheaper on input and six to seven times cheaper on output through December 31, 2026. The real gap is smaller because 3.8 Flash spends more thinking tokens per task, and both Flash prices double on January 1, 2027.
Why does Gemini 3.8 Flash cost more than the price implies? It is built on 3.7 Flash and made to reason harder, burning more output tokens per request. Those tokens bill at the output rate, so the cost per completed task can run well above the headline number.
Which should I use for a coding agent? Use Gemini 3.8 Flash for volume work behind a review gate, and escalate hard or high stakes tasks to Claude Opus 5. Measure token spend on your own prompts before committing to either as a default.
When this applies to your stack
If you are building an agent that runs at volume, the model choice is a cost model, not a benchmark contest. The right answer is usually a cheap default with a measured escalation path to a frontier model, and the only way to set the threshold correctly is to instrument token spend per task under your real traffic. That routing layer, and the evaluation harness that decides when to escalate, is the part that actually determines your bill. If you want a second set of eyes on how to structure model routing and cost controls for an agentic workload, that is the kind of AI integration and agent engineering work we do at Contra Collective.
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.