Gemini 3.8 Flash: Stronger Rows and a Dated Offer
MODELS & COSTS
Gemini 3.8 Flash: Stronger Rows and a Dated Offer
Some of Google's launch rows make 3.8 Flash a useful candidate for a low-cost trial. Other rows still favor larger models. The introductory rate also expires, so a continuing workflow needs both a matched task test and a budget at the scheduled post-promotion price.
Gemini 3.8 Flash and 3.8 Flash Cyber — the reasoning-and-coding push, three weeks after 3.7. (image: Google)
The stronger rows at the introductory price
Intro pricing holds at $0.75 input / $3.75 output per million tokens through December 31 (then $1.50/$7.50, same pattern as 3.7). Claude Opus 5 costs $5/$25. GPT-5.6 Sol costs $4/$20. So the Flash is six to seven times cheaper than the flagship it's now beating on real workloads — and Google's own published table shows wins on finance, legal, video, biology, and chart reasoning.
Where the $0.75 model wins. Vals Finance 61.4 vs Opus 5's 58.6; Harvey Legal 10.0 vs GPT-5.6 Sol's 2.5 (four times); LVBench 87.8 vs Opus 5's 75.4. Vendor-reported, day-1, redrawn.
The legal number deserves a second look. Harvey's Legal Agent Benchmark — complex legal workflows — shows Flash at a 10.0% pass rate versus GPT-5.6 Sol's 2.5% and Opus 5's 6.7%. Four times the OpenAI flagship at a sixth of the price. Finance (Vals Agent v2) and long-video understanding (LVBench, 87.8 vs 75.4) are similarly decisive. This is Google going after the workflows enterprises actually bill for — not just the coding leaderboards.
Tasks where Flash still trails
Google isn't claiming an outright win, and neither should we. The flagship gap shows up exactly where you'd expect — general agentic ability and long-horizon work. The standout is Terminal-bench 4.0: 19.1% for Flash vs 51.8% for Opus 5. That's not a close race; that's a different class of general-purpose agent. OSWorld computer use (59.0 vs 75.4), knowledge-work Elo (1545 vs 1824 on GDPval), and DeepSWE long-horizon software engineering (71.0 vs 74.0) all still favor Opus 5.
The honest read: on general agentic ability, the flagship is still the flagship. Terminal-bench 4.0's 51.8 vs 19.1 is where a cheap fast model and a frontier reasoning model diverge most. Vendor-reported, redrawn.
Against its own predecessor, though, 3.8 Flash is better almost everywhere: Terminal-bench 4.0 up from 11.2, OSWorld up from 50.6, BioMystery's hard tier up from 43.5 to 56.5. The model card describes it as built on 3.7 Flash with gains in software engineering and agentic knowledge workflows — an iteration, not a reinvention. Same 1M context, same 64K output, effort levels intact.
Google's own eval table — price on top, wins shaded. Flash leads most rows but Opus 5 takes GDPval, Terminal-bench 4.0, and OSWorld. (image: Google, via 9to5Google)
The rate and its expiry
Look at the chart below and the pattern is unmistakable. Google isn't trying to out-frontier Anthropic at the top of the market. It's pricing Flash so far below the flagships — while getting within striking distance on the workflows that matter — that for any business running agents at scale, the decision stops being about benchmarks and becomes arithmetic.
Six to seven times cheaper per token than Opus 5. Intro pricing holds to Dec 31 2026, then $1.50/$7.50. Vendor list prices, redrawn.
The finance, legal, video and biology rows suggest tasks to evaluate, not a general reason to replace a stronger model in those domains. The cited terminal benchmark still favors the flagship. Choose a representative task and compare accepted results at the effort and budget the application will use.
The chart is Google’s launch-day table, and rival scores may come from earlier snapshots. Keep that date distinction in the comparison. For budgeting, the cited introductory rate ends December 31, with $1.50/$7.50 listed afterward.
Sources: Google announcement (X) · DeepMind model card · 9to5Google · officechai · benchmarks vendor-reported as of September 2, 2026. Related: my Gemini 3.7 Flash post.
New flagship: Gemini 4 Argon: 1M Output Tokens, $2/$10 API Pricing.
Comments
Post a Comment