Google Gemini 3.8 Flash vs Opus 5 vs GPT 5.6

Google's Gemini 3.8 Flash landed on 2 September 2026 with benchmark charts placing it above Claude Opus 5 and GPT-5.6 — while costing roughly a sixth as much per token.

That claim deserves scrutiny rather than a headline, because the benchmarks come from Google, on Google's page, chosen by Google. So this comparison does two things: reports the numbers exactly as published, then examines what they actually support. The short version is that one of the three headline results is a statistical tie, and the real difference between these models is not the scores at all.

The Three Models

Note the modality asymmetry: Gemini 3.8 Flash is the only one of the three that ingests video and audio natively. If your input is a meeting recording or a screen capture, this comparison is already over.

Price: The Real Gap

Per million tokens, USD:

Two things the headline framing hides. First, Gemini Flash is not the cheapest model here — GPT-5.6 Luna undercuts it nearly fourfold. Second, Flash's advantage halves in January when introductory pricing ends.

GPT-5.6 also charges more above a context threshold (Sol rises to $8/$30 on long context), and both OpenAI and Anthropic discount heavily for cached input — Sol's cached input is $0.40, and Opus 5 cache reads run at a tenth of base input price. On a workload with a large stable prefix, the headline gap narrows considerably.

The Benchmarks, As Published

These are Google's figures from its own DeepMind pages, reproduced exactly.

HLE-Verified — expert multidisciplinary reasoning

The top three are separated by 0.5 percentage points. On a benchmark of this size that is a tie, not a ranking — run it again on a different day and the order could reshuffle. Reporting "Gemini beats Opus 5 and GPT-5.6 on HLE" is technically true and practically meaningless.

What is meaningful: a model at $0.75/$3.75 landed in a dead heat with models at $4/$20 and $5/$25. The story is not that Flash won. It is that the price of that score fell by roughly 85%.

Vals Finance Agent v2

A 2.8-point lead over Opus 5 — narrow but outside tie territory. The oddity worth flagging: GPT-5.6 Terra outscores GPT-5.6 Sol here, despite Sol being the flagship at double the price. When a vendor's cheaper model beats its flagship on your benchmark, it is a sign the benchmark is measuring something specific rather than general capability.

Harvey Legal Agent

Read the axis before the ranking. The winning score is 10%. Every model here fails at least 90% of the tasks. Gemini leading by 3.3 points over Opus 5 is a 1.5× relative difference on a benchmark where the honest summary is "none of these models can do this job yet."

If you are evaluating AI for legal work, the actionable finding is not which model to pick — it is that this task class remains unsolved and needs a human in the loop regardless of vendor.

How Much to Trust These Numbers

Three caveats that apply to every vendor benchmark chart, this one included:

• Google selected the benchmarks. A vendor publishes the evaluations it does well on. The absence of GPQA, AIME, MMMU, LiveCodeBench and standard SWE-bench from this comparison is information in itself.
• Configuration is rarely disclosed. Reasoning effort, scaffolding and retry policy move agentic scores substantially. Opus 5 runs five effort levels and GPT-5.6 six — a comparison against them at low effort is a different result than at max, and the settings are not published.
• Google's own charts are not a clean sweep, which is a point in their favour. On CWE-Bench patching, Google reports its Flash Cyber variant at 47.2% against a leading frontier model's 47.8% — a published loss. That is a more credible presentation than uniform victory.

The reasonable posture: treat these as evidence that Flash is competitive with the frontier tier on some agentic tasks, not that it surpasses it. Then run your own eval, because the only benchmark that predicts your results is your workload.

What Each Model Is Actually For

Gemini 3.8 Flash

The price-performance pick, and the only one that ingests video and audio. Strongest fit: high-volume agentic pipelines, document and media processing, finance-style analysis, anything where per-token cost is the binding constraint. Weakest fit: work needing more than 64K of output, or the deepest reasoning settings — its three thinking levels top out below what Opus 5 and GPT-5.6 offer.

Claude Opus 5

The frontier option when correctness outranks cost. Five effort levels up to max, 128K output, adaptive thinking on by default, and the deepest agentic tooling of the three — task budgets, compaction, context editing, mid-conversation system messages. Strongest fit: long-horizon coding and agent work where a wrong answer is expensive — the workloads our Opus 5 prompt collection is built around. Weakest fit: high-volume routine calls, where you pay 6.7× for capability you are not exercising.

GPT-5.6

The one that is really three models, and the variant choice matters more than the family. Luna at $0.20/$1.20 is the cheapest serious model in this comparison and belongs in any high-volume cost calculation. Terra at $2/$12 is the balanced middle — and outscored Sol on Vals Finance. Sol is the flagship for demanding professional work. Six reasoning levels including none gives the finest-grained cost control of the three.

Which Should You Choose?

The Verdict

The headline is wrong and the underlying story is bigger than the headline. Gemini 3.8 Flash did not decisively beat Claude Opus 5 and GPT-5.6 — on HLE the three are separated by half a point, which is noise. On Google's two agentic benchmarks Flash leads more clearly, but they are Google's chosen benchmarks and one of them has a 10% ceiling.

What is genuinely notable is the price at which Flash reaches that dead heat. A workhorse model matching frontier scores at a sixth of the per-token cost changes the default: the burden of proof has shifted onto the expensive models. You should now be able to say why a task needs Opus 5 or Sol rather than assuming it does.

There are good answers to that question — 128K output, the deepest reasoning settings, Opus 5's agentic tooling, and the reliability that matters when errors are costly. But "it is the best model" is no longer one of them by default, and that is a real change from six months ago, when Opus 5 comfortably outclassed Gemini 3.6 Flash.

Two practical notes to close on. Model your 2027 costs at Flash's post-January rate. And if you have a large cached prefix, redo the arithmetic with cache pricing before concluding anything — it moves the answer more than any benchmark on this page.

Keep Reading

If you are weighing the subscription route rather than the API, Claude pricing explained covers Pro, Max, Team and Enterprise. Or browse all guides and prompts on PromptsRush.