Gemini 3.8 Flash vs 3.7 Flash: 3 Regressions

Google shipped gemini-3.8-flash to general availability on September 2, 2026 — twenty days after gemini-3.7-flash went GA on August 13. If you built on 3.7 Flash three weeks ago, you have exactly one question: is this worth touching production for?

The unusual thing about this pair is that there is nothing to compare on paper. Same price, same context window, same output ceiling, same modalities, same parameters — every dimension a spec sheet normally settles is a tie. That makes it a pure quality-delta question, and the honest answer is narrower than the launch coverage claimed — including the first version of this page. Independent measurement now puts the gain at under two points of Intelligence Index, bought with three regressions: slower streaming, a slower first token, and 33% more cost per completed task.

Updated 5 October 2026. Artificial Analysis rebased its Intelligence Index from v4.1.1 to v4.3.2 in October 2026 and published no conversion factor, so every independent score below is restated against v4.3.2 and stamped with that version. Two composites this comparison previously leaned on — the standalone Coding Index and the Agentic Index — were retired in the rebase and have no v4.3.2 equivalent. Those rows are deleted rather than converted, and the verdict that rested on them has changed.

What actually changed between Gemini 3.8 Flash and 3.7 Flash?

On the published specification, nothing. Both model pages on Google's Gemini API docs list identical values, and the OpenRouter models API (machine-read September 3, 2026) returns the same figures independently. The identical rows are the finding, so here they are in full:

One pricing note applies to both equally: the $0.75/$3.75 rates are introductory through December 31, 2026. On January 1, 2027 both models move to $1.50 input and $7.50 output. Upgrading does not reset that clock, and staying on 3.7 does not protect you from it.

What does Google say improved in Gemini 3.8 Flash?

Google's framing is narrow and, to its credit, honest about scope. The API changelog describes 3.8 Flash as "our most intelligent Flash model, engineered for long-horizon software engineering, autonomous agents, and complex enterprise workflows," and reclassifies 3.7 Flash as "our previous-generation Flash model." The DeepMind Flash model page publishes three head-to-head numbers:

Notice what Google chose to publish: two agent benchmarks and one expert-reasoning benchmark. There is no head-to-head coding table, no MMLU-style knowledge comparison, no speed claim. When a vendor picks its own three slides and the largest gap is 2.4 points, that is a signal about where the model actually moved.

The page also carries a customer testimonial claiming 3.8 Flash completed "more than three times as many tasks as Gemini 3.7 Flash" on document-heavy workflows. That is an unaudited quote from one evaluation on one workload, not a benchmark — and no independent source reproduces it.

How do the two compare on independent benchmarks?

Vendor slides only matter if a third party running both models through the same harness agrees. Two do.

Artificial Analysis runs both at high reasoning effort through the same ten-evaluation suite. Here are the v4.3.2 figures, captured 5 October 2026 — and they do not say what the v4.1.1 figures said:

That table is the whole article in five rows, and four of them are not what an upgrade is supposed to look like. The quality gain is real but small: 1.87 points of Intelligence Index v4.3.2, which moves Gemini 3.8 Flash from rank #61 to rank #50 in a 224-model field — mid-table to slightly-less-mid-table, not a tier change. Against that single gain sit three measured regressions:

• Throughput fell. 279.7 tokens per second down to 239.1, roughly 15% slower. The older model is not merely faster, it is the fastest model Artificial Analysis currently measures — rank #1 of 224 on output speed. Giving that up is a real cost.
• Time to first token got worse by 2.83 seconds, 10.62s to 13.45s. That is a 27% increase in dead air before anything appears, and it is the direct consequence of reasoning harder by default.
• Cost per completed task rose 33%, $0.93 to $1.24, on an identical rate card. Thinking tokens bill at the output rate, so a model that deliberates more costs more at the same price per token.

Two rows that used to anchor this comparison are gone. Artificial Analysis's standalone Coding Index and Agentic Index were retired when the composite rebased to v4.3.2 — there is no v4.3.2 successor to either, and no conversion. This page previously built its entire recommendation on a 4.9-point Agentic Index gain (50.0 against 45.1, both v4.1.1-era figures). That measurement no longer exists, so the recommendation built on it does not survive either. What v4.3.2 does carry is an Agents block weighted at 30% of the composite (AA-Briefcase v1.1, GDPval-AA v2.1, AutomationBench-AA), alongside General 30%, Coding 20% and Scientific Reasoning 20% — and the all-in composite result of that is the +1.87 above.

LMArena (now served at arena.ai) is the one independent axis where 3.8 Flash looks clearly better, and with a month more data it has separated. As of 5 October 2026, Gemini 3.8 Flash (High) sits at rank 8 with 1495 ±5 Elo on 26,298 votes; Gemini 3.7 Flash (High) sits at rank 17 with 1488 ±5 on 21,539 votes. Seven points with ±5 bars on both sides still leaves the intervals touching, and both scores remain flagged preliminary, so read it as directional rather than settled. Nine ranks of separation on a blind-preference board is the strongest case for 3.8 Flash on this page — and it is worth weighting precisely because nobody can rebase it. For scale, the top of the board is Gemini 4 Argon (High) at 1525 ±9.

Design Arena is the one place the older model is ahead. On the website-generation category, 3.7 Flash holds 1318 Elo at a 57.1% win rate (rank 6) against 3.8 Flash's 1311 Elo at 56.3% (rank 11). Those are still the standing figures on OpenRouter's mirror of the board a month after launch, so this is no longer a one-day-of-data artefact — but a seven-point Elo gap on a taste benchmark is suggestive, not decisive. What it is not is evidence of an upgrade.

Is Gemini 3.8 Flash cheaper or more expensive to run?

Per token, identical. Per task, meaningfully more expensive — and this is the part teams get wrong.

Both models bill internal reasoning tokens at the $3.75/1M output rate. Google's pricing page states outright that output pricing "includes thinking tokens," and OpenRouter lists an explicit internal_reasoning price of $3.75/1M on both. Thinking is mandatory on both models and defaults to medium effort; minimal returns an error. You cannot opt out of paying for reasoning, only turn it down.

That matters because Artificial Analysis bills what it actually spends. Running the v4.3.2 Intelligence Index, Gemini 3.8 Flash cost $1.24 per task against 3.7 Flash's $0.93 — 33% more, at an identical per-token rate card, purely because 3.8 Flash thinks longer per task. Compare per-task figures rather than whole-suite totals here: Artificial Analysis runs different task counts per model, so suite totals are not directly comparable between the two, and the per-task column is the one that maps onto your invoice.

There is a sharper way to feel this. The closest open-weight rival in the bracket beats Gemini 3.8 Flash on both axes that matter to a queue:

GLM-5.3-Flash scores higher than either Gemini on the same index at roughly a fifth of Gemini 3.8 Flash's cost per task. What Gemini keeps is throughput — 3.8 Flash streams about 4.6× faster, and 3.7 Flash faster still. Behind a user-facing stream that is decisive; in an overnight queue it is close to worthless.

So the practical read: expect 3.8 Flash to cost about 1.33× per completed task at high reasoning effort, for 1.87 points of Intelligence Index v4.3.2. On a $200/month workload that is an extra $66 for a difference your users are unlikely to notice. On an agent that was failing a third of its long-horizon runs and now completes them, the same multiplier is trivially worth paying — but that is a claim you have to verify on your own traffic, because the independent composite no longer isolates agentic performance for you.

Reasoning effort is the lever, and on this pair it behaves oddly. Dropping either model to low does not simply trade quality for money: Artificial Analysis measures Gemini 3.8 Flash (Low) at 33.45 index and $1.47 per task — lower quality and higher cost than the High tier — and 3.7 Flash (Low) at 36.95 index and $3.46 per task, nearly four times the High tier's cost. The sweet spot on both is Medium: 3.8 Flash (Medium) at 39.77 for $0.93, and 3.7 Flash (Medium) at 39.62 for $0.29. Read those four numbers again, because they contain the real finding of this comparison: 3.7 Flash at Medium effort scores within 0.15 index points of 3.8 Flash at Medium effort, for less than a third of the cost per task. If you only change one setting after reading this page, make it thinking_level, not the model string.

Is Gemini 3.7 Flash deprecated?

No. As of September 3, 2026, Google's deprecations page lists no announced shutdown date for gemini-3.7-flash, and it remains listed as a Stable model in the API docs, not Legacy. The only Flash-family shutdown currently on the calendar is gemini-3.1-flash-lite on May 7, 2027.

Google's own wording on that page is worth internalising: published shutdown dates "indicate the earliest possible dates on which a model might be retired," with advance notice before retirement. Since 3.7 Flash has no date at all, you are not on a clock. Nothing about the 3.8 release forces a migration — it reclassifies 3.7 as previous-generation, which is a marketing state, not an operational one.

The real deadline on your calendar is not deprecation, it is January 1, 2027, when both models double in price on every tier at once: standard to $1.50/$7.50 per 1M, batch and flex to $0.75/$3.75, priority to $2.70/$13.50, cached input reads to $0.15 per 1M, and cache storage to $1.00 per 1M tokens per hour. The discounts double alongside the list price, so the relative saving from batching and caching survives intact — the absolute bill does not. Budget for that, not for a forced upgrade.

What do you need to change in your code to switch?

The model string. That is genuinely it, and we checked each surface that could have broken:

• Parameters: OpenRouter reports an identical supported-parameter set for both — temperature, top_p, seed, stop, max_tokens, tools, tool_choice, response_format, structured_outputs, reasoning, reasoning_effort, include_reasoning. No additions, no removals.
• Thinking config: both accept low/medium/high, both default to medium, both reject minimal. Existing thinking_level code carries over unchanged.
• Limits: identical 1,048,576 input and 65,536 output, so no prompt-chunking or truncation logic needs revisiting.
• Modalities and tools: identical input types and identical tool support, so multimodal pipelines and function-calling schemas transfer as-is.

In the Google GenAI SDK that is one string: model='gemini-3.7-flash' becomes model='gemini-3.8-flash'. On OpenRouter, google/gemini-3.7-flash becomes google/gemini-3.8-flash.

What is not a no-op is behaviour. A model that deliberates longer will change your latency profile — time to first token is 2.83 seconds worse and streaming is about 15% slower — along with your per-request token budgets and any prompt you tuned against 3.7's more concise style. If you have output-length caps, retry logic on max_tokens, or a hard per-request cost ceiling, re-check those before rolling out. Ship it behind the same flag you would use for any model change and watch cost per completed task, not cost per token.

Who should upgrade, and who should stay on 3.7 Flash?

The default is now: stay on Gemini 3.7 Flash. That is a reversal of how this release was first reported here and almost everywhere else, and it follows directly from the rebased measurements. Three of the four independent metrics moved against the newer model, and the fourth moved by less than two points.

Stay on Gemini 3.7 Flash if you are latency-sensitive in any way. 3.8 Flash is 2.83 seconds slower to first token and streams 15% slower, and 3.7 Flash is the single fastest model in Artificial Analysis's 224-model field on output speed. Behind a streaming chat UI, a voice front-end, or an autocomplete surface, this release is a plain regression with nothing offsetting it.

Stay on Gemini 3.7 Flash if you are cost-sensitive at acceptable quality — which covers high-volume chat, classification, extraction, summarisation and straightforward code generation. You would pay 33% more per completed task for 1.87 index points. And if you have not tuned thinking_level, that is where your money is: 3.7 Flash at Medium effort scores 39.62 for $0.29 per task against 3.8 Flash at Medium's 39.77 for $0.93.

Switch to Gemini 3.8 Flash if your workload is long-horizon, agentic and asynchronous — chained tool calls, autonomous software-engineering loops, document-heavy enterprise pipelines where a task either completes or doesn't and nobody is watching a cursor blink. All three benchmarks Google chose to publish are agent benchmarks, 3.8 Flash wins all three, and a 2.83-second first-token penalty costs a background job nothing. If a failed run costs you more than a 1.33× bill, upgrade.

Price GLM-5.3-Flash instead if the job is batchable. It scores 41.81 on the same v4.3.2 index — above both Gemini models — at $0.25 per task. The only thing you surrender is speed, and in a queue speed is nearly free.

Test both if your workload sits in between: RAG pipelines with light tool use, multi-step but short-horizon flows. Run 200 real requests through each and measure cost per successfully completed task plus p95 wall-clock, not index points. Since the independent composite no longer isolates agentic ability, your own traffic is the only harness that can settle this.

For the wider Gemini picture, our Gemini 3.7 Flash launch guide covers the pricing structure and benchmark provenance in more depth, the Gemini 3.8 Flash complete guide goes deeper on the new model on its own terms, and the Gemini 3.5 guide is the right reference if you are still on the May 2026 generation. If cost per task is what's driving this decision, it is also worth benchmarking against GLM-5.3 Flash before you commit to either Google model.

So what is the honest verdict?

This is a point release that bought a small quality gain with three measurable regressions, and it was mis-sold — by Google's framing, by the September coverage, and by the first version of this page. Gemini 3.8 Flash scores 1.87 points higher on Intelligence Index v4.3.2, wins the three agent benchmarks Google published, and leads on blind human preference. It is also slower to start, slower to stream, and 33% dearer per completed task, and the composite that used to justify the upgrade on agentic grounds no longer exists.

The decision rule: if your failures are agents giving up halfway through a long asynchronous task, upgrade. If your failures are anything else — latency, cost, coding throughput — this release makes two of the three worse. Stay on 3.7 Flash, move it to Medium thinking effort, and spend the saving on evaluating GLM-5.3-Flash before January, when both Gemini prices double.

FAQ

What's the difference between Gemini 3.8 Flash and 3.7 Flash?

On specifications, nothing — same $0.75/$3.75 pricing, same 1M context, same 65,536-token output, same modalities and parameters. The difference is quality traded against speed and cost. On Artificial Analysis's Intelligence Index v4.3.2 (October 2026) 3.8 Flash scores 40.93 against 3.7 Flash's 39.06, but it streams 40.6 tokens per second slower, takes 2.83 seconds longer to first token, and costs $1.24 per index task against $0.93.

Is Gemini 3.8 Flash more expensive?

Not per token — both are $0.75 input and $3.75 output per 1M through December 31, 2026, rising to $1.50/$7.50 on January 1, 2027. Per task it is 33% more expensive, because 3.8 Flash emits more thinking tokens and those bill at the output rate. Artificial Analysis measures $1.24 per Intelligence Index v4.3.2 task on 3.8 Flash against $0.93 on 3.7 Flash.

Should I upgrade to Gemini 3.8 Flash?

Usually not. Upgrade only if you run long-horizon agents, autonomous coding loops, or document-heavy enterprise workflows that run asynchronously — 3.8 Flash wins all three agent benchmarks Google published, and the extra latency costs a background job nothing. Stay on 3.7 Flash for anything latency-sensitive or cost-sensitive: you would pay 33% more per completed task for 1.87 points of Intelligence Index v4.3.2, while giving up the fastest output speed Artificial Analysis currently measures.

Is Gemini 3.7 Flash deprecated?

No. As of September 3, 2026, Google's deprecations page lists no shutdown date for gemini-3.7-flash, and the API docs still classify it as Stable. Google reclassified it as "previous-generation" in its marketing copy, which carries no operational consequence. The only date that affects it is January 1, 2027, when introductory pricing ends for both models.

Do I need to change my code?

Only the model string: gemini-3.7-flash to gemini-3.8-flash. Parameters, thinking levels, token limits, tool support and modalities are identical, so nothing else breaks. Do re-check output caps, latency budgets and cost ceilings, because 3.8 Flash generates substantially more reasoning tokens and has a slightly higher time-to-first-token.

Which is better for coding?

There is no longer a clean independent answer. Artificial Analysis retired its standalone Coding Index when it rebased to Intelligence Index v4.3.2 in October 2026; coding now sits inside the composite as a 20% block (Terminal-Bench 4.0 and SciCode) and is not reported separately. The older v4.1.1 Coding Index had the two models within 0.2 points of each other, but that figure does not convert. On Design Arena's website-generation category 3.7 Flash is still marginally ahead (1318 against 1311 Elo). For pure code generation there is no measured reason to move, and two reasons not to: latency and cost per task.

Which is faster?

Gemini 3.7 Flash, on both axes. Artificial Analysis measures 3.7 Flash at 279.7 tokens per second against 3.8 Flash's 239.1 — the fastest output speed in its 224-model field — and a 10.62-second time to first token against 13.45 seconds. The older model starts sooner and streams faster, which is the opposite of what a Flash-tier refresh normally delivers.

Artificial Analysis figures captured 5 October 2026 against Intelligence Index v4.3.2; the rebase from v4.1.1 has no conversion factor, so do not compare them to scores published before October 2026. LMArena and pricing re-verified the same day. Other sources verified September 3, 2026: Gemini API changelog, Gemini API pricing, the 3.8 Flash and 3.7 Flash model pages, Gemini deprecations, DeepMind Gemini Flash, Artificial Analysis, LMArena, and the OpenRouter models API.