What changed on 2 September
- Another Flash, three weeks after the last one. Google released Gemini 3.8 Flash on 2 September 2026. By Artificial Analysis's count it is the company's fourth Flash model in under four months.
- The index moved three points. Artificial Analysis scores it 59 on its Intelligence Index at high reasoning, up from 56 for Gemini 3.7 Flash at the same setting.
- Fifty-nine is a frontier-tier number. The same 59 is recorded for GPT-5.6 Sol at xhigh reasoning and for Grok 4.6 at medium reasoning.
- The gain is agentic. Artificial Analysis attributes it primarily to tau-cubed-Banking (tool use), Terminal-Bench v2.1 (coding) and GDPval-AA v2 (real-world tasks) rather than to the academic components.
- One published component figure. tau-cubed-Banking comes in at 45%, a twelve-point gain over Gemini 3.7 Flash.
- The window is unchanged. One million tokens of context, the same as 3.7 Flash.
- The price has an expiry date. Promotional API pricing of $0.75 input and $3.75 output per million tokens runs to the end of 2026. Standard pricing is $1.50 and $7.50, from 1 January 2027.
- Cost per task rose about 40%. Artificial Analysis measures $0.58 per Intelligence Index task at high reasoning, against $0.40 for Gemini 3.7 Flash. The per-token price is unchanged; the token count is not.
- Artificial Analysis puts it on the frontier. Not the intelligence frontier — the Intelligence versus Cost per Task Pareto frontier, which is a different and more useful claim.
The release itself is unremarkable. Google shipping a Flash model in September 2026 is roughly as newsworthy as Google shipping a Flash model in August, July or June, which it also did. What is worth twenty minutes of your week is the second bullet standing next to the third. Almost every production LLM stack built in the last two years rests on an assumption that has now come loose: that the cheap tier handles the easy work, and anything genuinely hard has to be escalated to a frontier model. On one independent evaluator's composite, the cheap tier is now scoring where the frontier tiers score.
One caution before any arithmetic. The Intelligence Index is a composite produced by a single evaluator, and a composite is a weighted opinion about which capabilities matter. It is a good signal — arguably the best public one, because Artificial Analysis runs the evaluations itself rather than reprinting vendor claims — but it is not your workload, and no index score has ever survived contact with a specific production distribution unchanged.
The same score, at different reasoning effort
Put the four numbers in one place and the interesting column is not the one on the right.
| Model | Reasoning effort | Intelligence Index |
|---|---|---|
| Gemini 3.8 Flash | High | 59 |
| Gemini 3.7 Flash | High | 56 |
| GPT-5.6 Sol | xhigh | 59 |
| Grok 4.6 | Medium | 59 |
GPT-5.6 Sol reaches 59 at xhigh — its most expensive setting. Gemini 3.8 Flash reaches the same 59 at high. Grok 4.6 gets there at medium. Two honest cautions apply to reading that column. Effort labels are each vendor's own and do not map onto a shared scale, so "medium" at one lab is not "medium" at another. And within a single vendor, a higher effort setting generally means more reasoning tokens, which means more money and more latency for the same finished answer.
The comparison that survives both cautions is the one Artificial Analysis draws for itself: cost per task. On that chart it places Gemini 3.8 Flash on the Intelligence versus Cost per Task Pareto frontier — meaning nothing currently evaluated delivers more intelligence for less money per task. That is a narrower claim than "as good as the frontier", and a more actionable one. It says a Flash-tier model is now on the efficient boundary of the whole field, not merely the best of the cheap options. What it does not say is that the model got cheaper: on Artificial Analysis's own measure, its cost per task went up rather than down, and the figures are in the pricing section below. We have seen the same shape from the other direction this year, in Grok 4.6 reaching parity with GPT-5.6 Sol at a fraction of the price and in Claude Sonnet 5 landing near Opus performance well below Opus pricing. The pattern is no longer an anomaly.
What the index actually measures
Version 4.1.1 of the Intelligence Index is a composite of nine evaluations: GDPval-AA v2, tau-cubed-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience and AA-LCR. Read that list as roughly four families — agentic and tool-use work, coding, hard scientific and reasoning problems, and knowledge plus long-context recall.
Nine components matter because of what averaging does. A three-point move on a nine-part composite is a small headline hiding a much larger movement somewhere inside it. Three points can mean nine components each drifting up slightly, or two components jumping and seven standing still. Those are entirely different products, and only the second one is a genuine capability change you can plan around.
The gain is agentic, not academic
In this case it is closer to the second. Artificial Analysis attributes the three-point gain primarily to the agentic evaluations — tau-cubed-Banking for tool use, Terminal-Bench v2.1 for coding, GDPval-AA v2 for real-world tasks — rather than to the science and knowledge components. The one component number published is the sharpest one: tau-cubed-Banking at 45%, twelve points above Gemini 3.7 Flash.
A twelve-point jump on a banking tool-use benchmark is worth understanding, because that evaluation family tests something specific and awkward: multi-turn interactions where the model must call tools correctly, respect policy constraints it was given earlier in the conversation, and not quietly invent a state it never observed. That is precisely the shape of the work a Bengaluru fintech building a servicing agent or a London insurance scale-up automating claims triage is actually doing. It is also the shape of work that small, cheap models have historically been worst at, which is why so many teams route agentic turns to a frontier model by default.
Now read 45% plainly, because the number deserves it. Forty-five per cent means the model fails more of these tasks than it completes. A twelve-point gain on an evaluation where you still lose the majority of runs is progress, not arrival. If your agent architecture depends on a high per-turn success rate over long trajectories, this release does not rescue it — the compounding still eats you. What it does change is the point at which escalating to a more expensive model starts paying for itself.
Two things not to do with this data. First, do not read a three-point composite gain as a general capability upgrade — it is concentrated in agentic evaluations, so a summarisation or classification workload may see no benefit whatsoever from moving. Second, be careful with any Terminal-Bench v2.1 figure you see quoted for this model: Artificial Analysis does not publish one in this write-up, and the numbers circulating in secondary coverage disagree with each other. The safe statement is that Terminal-Bench v2.1 contributed to the gain, without a figure attached.
The price cliff on 1 January 2027
The pricing is the part most likely to cost someone real money, and it is the part getting the least attention. Promotional rates run to the end of 2026. Standard rates are exactly double on both the input and the output line.
| Line | Promotional, to 31 December 2026 | Standard, from 1 January 2027 |
|---|---|---|
| Input, per 1M tokens | $0.75 | $1.50 |
| Output, per 1M tokens | $3.75 | $7.50 |
| Illustrative: 400M input + 40M output per month | $450 | $900 |
The multiplier is two, on every line, with no tapering — and on a like-for-like workload it is a floor rather than a ceiling. The $450 and $900 figures hold token volume constant, which measured behaviour does not: Artificial Analysis records output tokens per task rising 30% between 3.7 and 3.8 Flash, so the price change and the token growth compound on the same job of work. A team that builds its 2027 forecast in October on the back of September's invoice will be understating this cost centre by 100%. That is a large enough error to change a hiring decision at a seed-stage company in Bengaluru, and large enough to matter in a UK scale-up's board pack, where inference cost has become a line item that non-technical directors now ask about by name.
Hedge it properly, though. Promotional pricing can be extended, and vendors have form for extending it when a model is winning share. Equally it may lapse exactly as published. The disciplined approach is to plan against the standard price, treat any extension as upside, and put the change into the model as a date-triggered switch rather than a comment in a spreadsheet nobody reads. Our guide to LLM unit economics, cost per task and margin covers the mechanics of getting that into a forecast that survives contact with a finance team.
Worth saying explicitly: even at the standard $1.50 and $7.50, this remains well below what the frontier tiers it now matches on the index charge for a comparable workload. The price cliff does not undo the economics of the release. It changes the size of the win from very large to merely large, and it does so on a date you can already put in a calendar.
Cost per task went up, not down
Here is the figure the per-token table cannot show you, and it is the one that ought to travel with the index score. Artificial Analysis prices Gemini 3.8 Flash at $0.58 per Intelligence Index task at high reasoning. Gemini 3.7 Flash, on the same measure at the same setting, cost $0.40. Nothing on the price list moved between them. The model simply spends more to get to an answer, and Artificial Analysis attributes the roughly 40% increase to a 30% rise in output tokens per task, to 48k, together with more turns on the agentic evaluations.
That reorders the risks. The doubling is a known change on a published date, which makes it the easy one to plan for. Token inflation is neither announced nor dated, it arrives with a model upgrade you chose, and on the evidence of these two releases it is already running at a rate that would swallow a good caching win whole. A model that thinks for longer at the same price per token is not a saving, however the rate card reads.
The lever a builder actually controls here is reasoning effort, and Artificial Analysis publishes what it costs. For Gemini 3.8 Flash, cost per task falls to $0.41 at medium reasoning and $0.24 at low. The low setting is therefore materially cheaper per task than 3.7 Flash at high, and the medium setting lands roughly level with it. What those settings give up in index score is not published, which is exactly why the third step below matters: the 59 is a high-reasoning number, and so is the $0.58.
Run the doubling through your cost model before you run the migration. If a switch to 3.8 Flash only pays for itself at promotional rates, it is not a migration, it is a four-month discount you will have to unwind in January. The decision you want is the one that still holds at $1.50 and $7.50 — and if caching and compression are already in play, our cache, route and compress guide shows where those savings stack with a model change rather than double-counting against it.
A million tokens at Flash prices changes the long-context sum
The context window did not move: one million tokens, unchanged from 3.7 Flash. That looks like a non-story until you notice that the useful property of a context window was never its size. It is the price behind it. A million-token window is an architectural option only if you can afford to fill it repeatedly, and at input prices typical of frontier tiers most teams cannot, which is why so many long-context windows go mostly unused in production.
At $0.75 per million input tokens, filling a large fraction of that window on a request stops being an extravagance and becomes a design choice you can defend in a review. The long-context-versus-retrieval decision, which has usually resolved in favour of retrieval on cost grounds alone, now has to be re-argued on its merits. That is not the same as saying stuffing wins. Retrieval still buys you precision, auditability and a much smaller latency budget, and a large window emphatically does not guarantee reliable recall across all of it — which is exactly why the index carries AA-LCR as a separate long-context component rather than assuming the window size settles the question. Our long context versus RAG decision guide walks the trade-off properly, and the cost axis in it has just moved.
One regional note, kept qualitative because the honest version is qualitative. API list prices are set in US dollars, so a team in Chennai and a team in Manchester face the same headline rate. What differs is how that dollar-denominated cost sits against a cost base that is mostly local — salaries, rent, revenue. A doubling on a dollar line lands differently on a rupee-revenue business than on a sterling-revenue one, and differently again for anyone selling into both. Work it in the currency your revenue arrives in, not the one the invoice is written in.
Every article here is written by a Verified Builder. Want your name on the next one?
AI Tech Connect lists AI engineers, founders and researchers across India and the UK — and the people hiring browse it to find them. Adding your profile is free.
Become a Verified Builder →What a builder should actually do this week
Nothing above justifies a migration. All of it justifies a measurement, and the measurement is small enough to fit inside a week. Five steps, in order.
One: put 3.8 Flash in the cheap slot and re-run your existing suite. Not a new benchmark, not a vibe check against a handful of prompts — the evaluation set you already trust, unchanged, with the cheap tier of your cascade swapped. If you do not have such a suite, that is the actual finding of the week, and it matters more than which model you pick. Cost-aware evaluation is the discipline here; our quality-per-pound evaluation guide covers how to score a model on outcome and spend in the same run rather than in two disconnected spreadsheets.
Two: measure the escalation rate, not the accuracy. In a cascade, the number that decides your bill is what fraction of traffic gets promoted to the expensive model. A cheap tier that is three points better on an index but escalates at the same rate has saved you nothing. A cheap tier that escalates ten percentage points less often has changed your unit economics, even if its standalone accuracy looks similar. The routing cascade guide sets out how to instrument that properly.
Three: test at the reasoning setting you will actually deploy. The 59 is a high-reasoning number. If your latency budget means you run at a lower effort setting in production, you are not buying the 59, and the published figure tells you comparatively little about what you will get.
Four: price both scenarios per task. Promotional and standard, on the same evaluation run, on your real token mix. Two columns, one spreadsheet, half a morning.
Five: do not rip out the frontier tier. Keep it as the escalation target. The claim on the table is that fewer requests need to reach it, not that none do — and on a tool-use evaluation where the new model completes 45% of tasks, an escalation path is not optional.
Whatever the comparison shows, write it up. A short, specific post — the suite you ran, the escalation rate before and after, the cost per task at both price points — is worth more to a hiring manager than any list of models you have used, because it is evidence of judgement rather than exposure. If you keep a Verified Builder profile, that write-up belongs on it.
What this release does not settle
Four limits are worth holding onto. The index is one evaluator's composite, and a good one, but a composite nonetheless. The cadence cuts both ways: four Flash models in under four months — we covered 3.6 Flash reaching general availability in July and 3.7 Flash arriving three weeks later at half the price — means the model you benchmark this week may not be the one you are deploying against by December, so build your evaluation harness to survive model churn rather than treating any single comparison as durable. The headline price is promotional, lapses on a published date, and is quoted per token rather than per task — measured cost per task moved the other way. And 45% on the agentic evaluation that improved most is a majority of runs still failing.
The honest summary is narrower than the index score suggests, and more useful than it. The cheap slot in a routing cascade has become good enough to be worth re-testing, on evidence from an independent evaluator, at a per-token price that is temporarily very low and permanently competitive — while the same evaluator measures it spending more tokens, and more money, on each task than the model it replaces. It has not become a replacement for a frontier tier, and anyone telling you otherwise this week is reading a single composite number and skipping the column next to it. Re-run your suite. Let the escalation rate decide.