Evidence note: This article rests on three kinds of material, and they are not equal. The first is primary vendor publication — Google DeepMind’s and Google’s own announcement posts and the published model card — which establishes what Google says it shipped. The second is a third-party aggregator leaderboard, read on a single date, which is a measurement someone else ran and not a vendor claim. The third is a scattering of public repositories, forum threads and one press report, which establish only that the artefacts exist, not that their contents are accurate. Each specific number below carries its own evidence marker. Nothing here was measured by this site.
Google shipped two models, not one
The framing that has followed this release — a Flash model, a leaderboard slot, a price — misses the shape of the announcement. Google DeepMind introduced Gemini 3.8 Flash and a second variant, 3.8 Flash Cyber, and Google’s own models-and-research channel published the release alongside it. Two artefacts, announced together, under one version number.
That matters downstream. Almost every comparison, ranking and cost figure circulating about this release attaches to the base Flash model. The Cyber variant is announced but largely absent from the third-party record examined here — no leaderboard row, no proxy identifier, no head-to-head repository. A reader who sees "Gemini 3.8 Flash scored X" should not assume the sibling model inherits the number.
The official record, and where it thins out
What Google has published on its own authority is a model card, available both as a model-card page and as a downloadable PDF. A model card is the strongest form of evidence in this set: it is the vendor committing in writing to a description of its own system.
Past that, the record thins. A developer documentation page for the model on Google’s AI developer site appears in the source set, which would indicate general Gemini API availability — but that availability has not been independently confirmed here, and the scope of the rollout (regions, tiers, quota) is open (evidence: single vendor documentation URL, unverified contents). Separately, a thread on the Gemini subreddit reports the model appearing in the Gemini web interface (evidence: user report on a public forum; no vendor confirmation of a web-app rollout was located, and whether it reached all users or a staged subset is open). A Hacker News discussion thread captures developer reaction but does not establish availability either way.
So: the model card is solid. API availability and web-app rollout are reported, not confirmed.
Eighth place, at 59
As the Artificial Analysis page for Gemini 3.8 Flash read on 22 September 2026, the model sat in eighth position with an intelligence index score of 59 (evidence: single third-party aggregator, one reading, one date; not reproduced by this site).
Three properties of that figure limit what it can carry. It is a composite — an index built by combining several underlying evaluations, so a single scalar hides which components moved. It is dated: leaderboards re-rank as new models land and as the aggregator revises its own methodology, and a rank read on 22 September 2026 is a claim about 22 September 2026. And it is one measurer. Nothing in the record examined here corroborates the score from a second source.
The one-point gap
On the same reading, Kimi K3 and GLM-5.3 both scored 60 — one point above Gemini 3.8 Flash, and ranked above it (evidence: same single aggregator reading, 22 September 2026).
A one-point difference on a composite index of this kind establishes approximate parity, not superiority. Composite scores of this construction typically have run-to-run variance from sampling temperature, prompt formatting and evaluation harness details; the aggregator does not publish, in the material reviewed here, a confidence interval that would tell a reader whether 59 and 60 are distinguishable at all. Treating the ordering as a ranking of capability reads more into the number than the number contains.
There is a concrete illustration of why single-number gaps mislead in our analysis of what 904 DeepSWE rollouts revealed about GLM-5.3 and Claude Fable 5 — two models that tied on the first attempt and separated only once the sample deepened. The lesson transfers: a headline score is a summary of a distribution, and summaries of distributions can agree while the distributions do not.
The column that inverts the ranking
The same comparison reportedly carries a cost-per-task column, and on that column the ordering reverses: Gemini 3.8 Flash at $0.58 per task against GLM-5.3 at $0.68 and Kimi K3 at $0.84 (evidence: reported figures from the same third-party aggregator, not independently verified; the task mix, token accounting and pricing tier behind the per-task figure are open).
If those numbers hold, the practical reading of the three-way comparison is close to the opposite of the rank order. A model one index point behind, at roughly 85 percent of the next-cheapest competitor’s cost per task and roughly 69 percent of the most expensive, is the economically dominant choice for any workload where a one-point composite difference does not bind. That is a derived comparison, not a measurement: the ratios are arithmetic on the reported figures, and they inherit every uncertainty those figures carry.
The caution worth stating plainly: cost per task is a function of the task set. A benchmark task mix that runs short favours models with cheap output tokens; a mix heavy in long agentic traces favours different economics entirely. Without the task definition, the $0.58 figure cannot be projected onto a reader’s own workload.
A configuration called ‘High’
The comparison figures are reported against a configuration labelled ‘High’ — a reasoning-effort or thinking-budget setting that third-party comparison tools and access wrappers select when they exercise the model (evidence: configuration label appears in third-party tooling; its exact semantics are not documented in the vendor material reviewed here).
This is not a footnote. If ‘High’ denotes an elevated reasoning budget, then both the intelligence score and the cost per task are measurements of the model under that budget, and neither transfers to a default-configuration deployment. A reader reproducing either number needs to match the configuration, and the configuration’s definition is currently open.
Reachable where Google did not put it
A distinct cluster in the public record concerns access paths. Google Antigravity is named as an environment where the model is reachable, including through third-party OAuth plugins and access workarounds — a claim that surfaces through community repositories such as an Antigravity access project and an OAuth authentication bridge for a third-party coding agent rather than through independent confirmation (evidence: third-party repositories; the vendor-sanctioned scope of this access is open).
Alongside those, several proxy and API-compatibility projects appear to expose a gemini-3.8-flash model identifier — among them an API-compatibility server, a proxy project and a free-access API wrapper, with a further repository built around the model in the same category (evidence: repository names and stated purposes; the projects’ contents were not executed or audited here, and whether the identifier actually resolves to the vendor model is open).
That last clause is the important one. An identifier string in a proxy’s model list proves that someone wrote the string. It does not prove the endpoint behind it serves Gemini 3.8 Flash.
A copy of the instructions that came from nobody’s documentation
An extracted system prompt attributed to Gemini 3.8 Flash reportedly appears in a public repository that collects system-prompt leaks (evidence: third-party collection repository; attribution and fidelity of the extracted text are unverified, and no vendor confirmation exists).
Extracted prompts are unusually weak evidence and unusually tempting to use. They arrive without provenance, they can be partially hallucinated by the very model being probed, and they may reflect a product surface’s wrapper rather than the model’s own conditioning. Treating such a document as a description of the model is a category error; at most it is a description of what one extraction attempt produced on one day.
Why the access layer changes how the scores should be read
Put the two clusters together and a methodological problem appears. The leaderboard numbers describe a model reached some particular way, under a configuration labelled ‘High’. The access layer described above reaches the same nominal model through proxies, plugins and wrappers whose routing, system-prompt injection and parameter defaults are not documented.
Two users can therefore both say "Gemini 3.8 Flash scored 59" and "Gemini 3.8 Flash performs worse than that for me" without either being wrong — because the identifier does not pin the pathway. Whether any of the third-party paths reproduce the aggregator’s numbers is entirely open; no source in this set tests it. This is the single largest gap in the public record around the release.
Used as a baseline elsewhere
A further signal, weaker but informative: the model appears to be functioning as a comparison baseline in independent benchmark projects. The source set includes a head-to-head repository pitting GPT-6 Astra against Gemini 3.8 Flash, a column-labelling race and a PDF extraction race (evidence: repository names and stated purposes; results, methodology, sample sizes and dates were not reviewed or reproduced here, and no conclusion from any of them is asserted).
What this establishes is adoption as a reference point, not performance. A model chosen as the baseline in labelling, document-extraction and agentic build comparisons is a model that independent developers consider the thing to beat — which is itself a fact about the release’s reception, and says nothing about whether it wins.
What the press framing leaves unmeasured
The Register covered the release as Google reminding everyone it is still in the race, dated 2 September 2026 (evidence: single press report; the characterisation is that outlet’s framing, not a measured finding).
The race framing is legible but not very useful to anyone deciding what to deploy. It compresses the interesting result — a model that ranks eighth and, if the cost figures hold, undercuts both models ranked above it by a wide margin — into a question about competitive standing. Standing is a story about vendors. Cost per unit of capability is a story about budgets, and it is the one with a decision attached.
Developer takeaway
What can be pinned down today, in descending order of confidence:
- Google announced two models under this version — a Flash model and a Cyber variant — and published a model card for the Flash model. Vendor-attested, from Google’s own channels. Conclusions drawn about the base model should not be extended to the Cyber variant without separate evidence.
- One third-party aggregator placed the model eighth at an index score of 59 on 22 September 2026, one point behind Kimi K3 and GLM-5.3 at 60. A one-point composite gap does not establish a capability difference.
- Reported cost per task of $0.58, against $0.68 and $0.84 for the two models ranked above it. Unverified, and dependent on an undisclosed task mix — but if it holds, it inverts the practical ranking.
- Availability on the developer API, in the web app, through Antigravity and through community proxies is reported rather than confirmed. Identifier strings in third-party tooling do not establish what answers behind them.
- A configuration labelled ‘High’ sits under the published figures, and its semantics are undocumented. Match it before comparing anything.
The question the record cannot answer, and which only a controlled run would settle: does the model reached through any of these third-party access paths, at a documented configuration, reproduce the aggregator’s intelligence score and cost per task on a task set the tester defines? A useful version of that experiment would fix the configuration, fix the task mix, run each access path with identical inputs, and report per-path token accounting alongside outputs — with enough samples to distinguish a real gap from the variance that makes one-point differences meaningless in the first place. That experiment has not been run here, and no source in this set has run it either.
One operational note worth carrying into any such plan: model availability is not permanent. Vendors and platforms retire endpoints on schedules that can arrive faster than a benchmark’s useful life, as the October 2026 retirement of selected Copilot models illustrates. Any cost-per-task advantage is a claim about a specific model at a specific price on a specific date — which is, in the end, the same caveat that attaches to the score.