Verdict: Gemini 3.8 Flash at high thinking effort wins on cost and turnaround; Claude Opus 4.6 stays the pick for long-horizon agentic persistence and very large single outputs. If you are asking which LLM is best for coding on cost and turnaround, the answer today is Gemini 3.8 Flash at high thinking effort. In our own controlled run, it matched Claude Opus 4.6 exactly on machine-checked instruction following while finishing in roughly a third of the wall time and costing a fraction as much per request. Opus 4.6 remains the pick when the work depends on long-horizon agentic persistence or a broader published benchmark record, but on the ordinary loop of planning, writing and revising code, the cheaper model no longer loses anything measurable.
TL;DR
- Across three trials each on an identical seven-constraint article-planning task, Gemini 3.8 Flash (High) and Claude Opus 4.6 (Thinking) both scored 17 of 17 on machine-checked constraint adherence. Median wall time was 23 seconds for Gemini against 67 seconds for Opus. (Our own measurement, n=6, three trials per model, measured 2026-09-16.)
- Gemini 3.8 Flash bills $0.75 per million input tokens and $3.75 per million output through 31 December 2026, then $1.50 and $7.50 from 1 January 2027 (Gemini API pricing).
- Claude Opus 4.6 bills $5.00 per million input tokens and $25.00 per million output (Anthropic pricing).
- Both models accept a roughly 1M-token context, so context size is no longer the deciding factor it was a year ago.
- Choose Gemini for cost- and latency-sensitive work; stay on Opus for persistence-heavy agent runs and very large single outputs.
- The 1 January 2027 price step-up on Gemini gives you a real deadline for locking in a decision.
Which LLM Is Best for Coding When You Measure It Yourself
Published leaderboards tell you how a model behaves on someone else's task. We wanted a check on ours, so we ran both models through the same harness with the same prompt: a seven-constraint article-planning task, scored programmatically rather than by human judgement, three trials per model via a headless CLI in print mode.
Across three trials each on an identical seven-constraint article-planning task, Gemini 3.8 Flash (High) and Claude Opus 4.6 (Thinking) both scored 17 of 17 on machine-checked constraint adherence. Median wall time was 23 seconds for Gemini against 67 seconds for Opus. That is our own measurement, n=6 across three trials per model, taken 2026-09-16.
Two things follow. First, at this task size the instruction-following gap between a Flash-tier model and a frontier Opus model is zero, not small. Second, Opus took about 2.9 times as long for the same result. In an interactive loop where you are reading the output and re-prompting, a 44-second difference per turn is the difference between a conversation and a wait.
The obvious limitation: n=6 on one task class is a signal, not a law. It tells you what happened on a bounded, well-specified planning job. It does not tell you what happens on a four-hour refactor across forty files.
What Does Each Model Actually Cost Per Request
Take a realistic single request: 100,000 input tokens of repository context and 10,000 output tokens of patch and explanation.
On Gemini 3.8 Flash at introductory rates of $0.75 input and $3.75 output per million tokens (Gemini API pricing), that is $0.1125. On Claude Opus 4.6 at $5.00 input and $25.00 output per million (Anthropic pricing), the same request is $0.75. Gemini comes out about 6.7 times cheaper.
The introductory window closes on 31 December 2026. From 1 January 2027, Gemini 3.8 Flash moves to $1.50 input and $7.50 output per million (Gemini API pricing), making that same request $0.225 and the gap about 3.3 times. Still decisive, but worth knowing the number changes.
Both vendors offer levers that move these figures. Anthropic prices cache hits at $0.50 per million tokens with five-minute cache writes at $6.25 and one-hour writes at $10.00, and discounts the Batch API by 50% (Anthropic pricing). Gemini supports caching plus batch, flex and priority consumption tiers (Gemini 3.8 Flash model page). If your workload has a stable system prompt and a fat repository preamble, caching narrows the gap more than either headline rate suggests. For a fuller breakdown of Anthropic's plan tiers against raw API billing, see our comparison of Claude Code pricing across Pro, Max and API.
When Is Claude Opus 4.6 Still the Right Pick
Three cases, and they are specific.
Long outputs. Opus 4.6 supports up to 128K max output tokens with a 1M context window (Opus 4.6 overview). Gemini 3.8 Flash accepts an input limit of 1,048,576 tokens but caps output at 65,536 (Gemini 3.8 Flash model page). If you are generating a large migration in one pass rather than chunking it, that ceiling matters.
Agentic persistence. Opus 4.6 uses adaptive thinking with a default effort of high, and it is the model Anthropic positions for long-running agent work. On multi-hour tasks where the model has to keep re-orienting after tool failures, the track record is longer.
Breadth of evidence. Opus has been benchmarked across more third-party evaluations for longer. If your organisation needs to justify a model choice against external results rather than internal ones, that history is an asset. Our real SWE benchmark comparison covers how the frontier tier stacks up on that basis.
What Do the Published Benchmarks Say About Gemini 3.8 Flash
Gemini 3.8 Flash shipped on 2 September 2026, described by Google as its best reasoning and coding model to date at the same speed and low cost as 3.7 (Introducing Gemini 3.8 Flash). In the same announcement, Google reports 54.9% on HLE-Verified, says the model outperforms most larger frontier models on DeepSWE v1.1 for long-horizon software engineering at a fraction of the cost, and claims wins on the Vals Finance Agent V2 and Harvey Legal Agent benchmarks.
Read those as vendor claims, because that is what they are. The useful detail buried in the announcement is behavioural rather than numeric: at higher effort the model takes extra reasoning steps and makes more iterative tool calls, spending more tokens on harder problems. That is why the High setting matters for coding, and why a low-effort comparison would understate it.
For how this generation compares against the other frontier options, see our verdicts on GPT-5.3 Codex versus Claude Opus 4.6, GPT-6 Astra versus Claude Fable 5.1, and the open-weight angle in GLM-5.2 versus Claude Opus 4.8.
How Should You Decide Before January 2027
Run your own six trials. The method costs an afternoon: pick one task your team actually does, write the constraints down so a script can check them, run three trials per model in a headless CLI, and record wall time. That is what produced the figures above, and it generalises better to your codebase than any leaderboard.
Then route by workload rather than picking one winner. Default to Gemini 3.8 Flash at High for planning, review, patch generation and anything interactive. Escalate to Opus 4.6 for long autonomous agent runs and single-pass outputs above 65K tokens. Most teams will find the escalation path handles well under a quarter of requests, which is where the cost saving actually lands.
FAQ
Q: Which LLM is best for coding in 2026?
A: For most day-to-day coding work, Gemini 3.8 Flash at high thinking effort. It matched Claude Opus 4.6 at 17 of 17 on machine-checked constraint adherence in our own n=6 run measured 2026-09-16, at about a third of the wall time and a fraction of the per-request cost.
Q: How much cheaper is Gemini 3.8 Flash than Claude Opus 4.6?
A: On a 100K-input, 10K-output request, Gemini costs $0.1125 against $0.75 for Opus at current published rates, roughly 6.7 times cheaper. From 1 January 2027 Gemini's rates double to $1.50 input and $7.50 output per million tokens, narrowing the gap to about 3.3 times.
Q: Does Gemini 3.8 Flash have a smaller context window than Opus 4.6?
A: No. Gemini 3.8 Flash accepts 1,048,576 input tokens and Opus 4.6 offers a 1M context window. The real difference is on output: 65,536 tokens for Gemini against 128K for Opus.
Q: Is a six-trial benchmark enough to choose a model?
A: Not on its own. Six trials on one task class shows whether a cheaper model is losing anything obvious on work you do repeatedly. It says nothing about multi-hour agentic runs, so treat it as a routing input rather than a verdict on all workloads.
Q: Should I use low or high thinking effort for coding?
A: High, for Gemini 3.8 Flash. Google notes the model spends more reasoning steps and tool calls on harder problems at higher effort, which is exactly the behaviour coding tasks need. Opus 4.6 defaults to high effort already.
Q: Do caching and batch discounts change the comparison?
A: They narrow it. Anthropic prices cache hits at $0.50 per million tokens and discounts batch requests by 50%; Gemini supports caching alongside batch, flex and priority tiers. If your prompts share a large stable prefix, measure with caching enabled before deciding.
Top comments (0)