Gemini 3.8 Flash Explained: Pricing Cliff, Context, Benchmarks

Gemini 3.8 Flash is Google’s September 2, 2026 mid-cycle Flash release: a 1,048,576-token-context, multimodal-input, text-output model priced at $0.75 per million input tokens and $3.75 per million output tokens until December 31, 2026. On January 1, 2027 those rates double. The headline says “same price as 3.7 Flash.” The invoice tells a different story.

That gap is the point of this article. Independent measurement from Artificial Analysis puts the model at $0.58 per Intelligence Index task at high reasoning, about 40% above Gemini 3.7 Flash’s $0.40, even though per-token prices are identical. The model “works harder”: more reasoning tokens, more tool calls, more turns. Per-token price is no longer a usable proxy for cost.

You will leave with the verified specifications, a benchmark table with caveats, a worked cost-per-task model that survives the January 2027 price cliff, a look at the gated Gemini 3.8 Flash Cyber variant, and a decision matrix for choosing between effort levels and competing models.

What this covers: lineage, architecture as far as it is disclosed, training (mostly undisclosed), benchmarks, access and pricing math, failure modes, and a comparison matrix.

Context and Background: Four Flash Models in Four Months

Google shipped its third Flash-tier release in about six weeks; CEO Sundar Pichai’s own phrasing was “Our 3rd Flash release in just 6 wks.” Artificial Analysis counts Gemini 3.8 Flash as the fourth Flash model in under four months. One reviewer tallied three Flash variants between July 21 and September 2, 2026. The counts differ by what you include, but the direction is unambiguous: Flash is now a rolling product line, not a yearly model.

For a builder, that cadence changes the operating model. You are no longer choosing “a model.” You are choosing a model string, a thinking level, and a price schedule, and any of the three can shift within a quarter. Our earlier deep-dives on Gemini 3.5 Flash and Gemini 3.5 Pro cover the earlier steps in this ladder; this post picks up at the point where Flash pricing dropped sharply and the economics got harder to read.

Price is where the lineage moved most. Google’s pricing page lists Gemini 3.5 Flash at $1.50 input and $9.00 output per million tokens. Gemini 3.7 Flash and 3.8 Flash both list $0.75 and $3.75 during the introductory window. That is 50% lower on input and roughly 58% lower on output than 3.5 Flash. After the cliff, input returns to $1.50 (equal to 3.5 Flash) and output lands at $7.50, still about 17% below 3.5 Flash’s $9.00.

Google’s positioning sits in the model card: the model is intended for “cost-effective scaling of general-purpose, production-ready agents,” especially software engineering, agentic tasks, and complex knowledge workflows. Read that sentence carefully. It describes a model tuned for long, tool-heavy loops, which is exactly the workload where token count, not token price, dominates spend. For the general cost landscape of running such loops, see our guide to AI inference cost optimization.

Primary sources for this article are Google’s launch post, the Gemini API model documentation, the Gemini API pricing page, and the DeepMind model card. Third-party numbers, chiefly Artificial Analysis, are labelled as such.

A note on conflicting figures

Secondary coverage disagrees in places. One summary of Google’s post quoted a $3.50 output price; Google’s pricing page and API docs both say $3.75, and so does the rest of the coverage, so this article uses $3.75. Artificial Analysis also publishes more than one Intelligence Index version, and a release-tracking page showed lower index values (41 for high) than the launch article (59 for high). This article uses the launch-day article, which is what the $0.58 cost-per-task figure is tied to, and flags the mismatch rather than mixing the two scales.

Gemini 3.8 Flash: What It Is and How It Is Built

Direct answer: Gemini 3.8 Flash (model ID gemini-3.8-flash, generally available) is a multimodal-input, text-output model with a 1,048,576-token input window and 65,536-token maximum output. Google says it builds on the Gemini 3.7 Flash architecture. Parameter count, expert layout, and training compute are not disclosed. Its differentiator is agentic behavior tuned for long coding and tool-use loops.

Figure 1: Gemini 3.8 Flash request path. Reasoning tokens are billed as output, and each tool call adds a new turn that re-reads the context.

Figure 1 shows the path that matters for cost. Multimodal input enters a 1M-token window. A thinking_level setting controls how much internal reasoning the model does. Tool results feed back in. Reasoning is billed as output. The loop repeats until the agent stops. Every arrow in that diagram is a place where tokens accumulate.

Disclosed specifications

The following are stated in Google’s documentation or model card:

Model ID: gemini-3.8-flash, generally available for production use.
Context window: 1M tokens (1,048,576 per third-party spec sheets); maximum output 64k (65,536).
Input modalities: text, images, audio, and video. Output is text.
Thinking levels: low, medium (default), and high. The minimal level is explicitly not supported.
Knowledge cutoff: the model card states training knowledge through March 2026, with some domains current only to earlier dates, consistent with the Gemini 3 family.
Architecture reference: the model card points to the Gemini 3.7 Flash card for architecture rather than describing a new one.

What is not disclosed matters as much. Google has not published a parameter count for Gemini 3.8 Flash, has not said whether it is a dense or mixture-of-experts network in the material we could verify, and has not disclosed its attention variant or tokenizer vocabulary size. Third-party blogs offer numbers for these; none cite a Google source, so we omit them. If a site gives you a precise parameter count for this model, treat it as speculation.

Thinking levels replace thinking budgets

The migration notes in the API documentation are the most concrete architectural signal available. Developers move from a numeric thinking_budget to a thinking_level string enum. The notes also list sampling parameters to remove (temperature, top_p, top_k, candidate_count) and say prefilled model turns are replaced by a server-side previous_interaction_id for conversation state.

Two consequences follow. First, you lose fine-grained control: three discrete levels instead of a token dial, so you cannot cap reasoning at, say, 2,000 tokens. Second, server-side interaction state means the provider, not your client, holds the conversation. That simplifies agent code and shifts the trust and retention questions to your data-governance review. It also affects caching design, covered below.

Why the 1M window is a cost feature, not a capacity feature

A 1,048,576-token window (roughly 1,500 pages by Artificial Analysis’s estimate) invites stuffing whole repositories or document sets into a prompt. Resist it by default. At $0.75 per million input tokens during the intro window, a full 1M-token prompt costs $0.75 uncached, and $0.075 when it hits the context cache. Post-cliff those become $1.50 and $0.15. A single full-window call is cheap; a 40-turn agent loop that resends most of it on every turn is not, unless the prefix is cached.

Our piece on context engineering for production LLM agents treats this in depth: the discipline is deciding what enters the window, not filling it.

Lineage: From 3.5 Flash to 3.8 Flash

The Flash line has moved fast enough that the version numbers hide the structure. Gemini 3.5 Flash carried $1.50/$9.00 pricing. A 3.6 Flash appeared next; Artificial Analysis references “Gemini 3.6 Flash (high)” as a comparison point. Gemini 3.7 Flash then reset the price floor to $0.75/$3.75 introductory. Gemini 3.8 Flash kept that price and raised capability.

Figure 2: Flash lineage and the Cyber branch. Price fell at 3.7 and held at 3.8; the behavior change at 3.8 is more reasoning and more tool turns.

What changed versus 3.7 Flash

Google’s own framing is behavioral: “3.8 Flash works harder. On complex tasks, it exhibits greater diligence, executing extra reasoning steps, and calling tools iteratively.” The launch post also says that for compute-efficiency-first work, developers can use lower effort levels “or continue to rely on Gemini 3.7 Flash, which remains fully supported for efficiency-first workloads.”

That is an unusual admission from a vendor. Google is telling you, in its own launch post, that the newer model is not automatically the cheaper way to finish a task.

The DeepMind model card adds two useful constraints. Automated testing shows 3.8 Flash performs “similarly to Gemini 3.7 Flash across both safety and tone, with low unjustified refusals.” And the Frontier Safety Framework assessment concluded the model “does not have meaningful new capabilities or material increases in performance” relative to 3.7 Flash from a frontier-risk standpoint. Read together with the benchmark gains, the picture is an incremental release: reported coverage describes the changes as “iterative rather than a full retraining run,” though that phrasing comes from a reviewer, not from Google.

Versus 3.5 Flash and 3.5 Pro

The question “Gemini 3.8 vs 3.5 Flash” has a clean answer on price and a murkier one on capability. On price, 3.8 is cheaper per token at every tier through 2026, and equal-or-cheaper afterward. On capability, Google did not publish a head-to-head table against 3.5 Flash in the launch post; it compared against 3.7 Flash and against larger frontier models. Artificial Analysis notes that 3.8 Flash at low reasoning (index 52) matches Gemini 3.6 Flash at high reasoning “at 30% lower cost.” Extrapolating from that, 3.8 at medium or high should sit above 3.5 Flash. We have not seen a published direct comparison, so treat that as an inference.

Against the Pro line, Flash trades peak capability for price and speed. Our Gemini 3.5 Pro explainer covers what the larger tier buys. The uncomfortable finding of 2026 is that a Flash model at high effort can beat older larger models on specific agentic benchmarks, which compresses the case for defaulting to Pro.

Training: What Google Has and Has Not Said

Training is the section where a deep-dive most often tempts writers into invention. The verified record is short.

Disclosed or directly implied by Google:

Gemini 3.8 Flash builds on the Gemini 3.7 Flash foundation; the model card defers architecture and training-data details to the 3.7 Flash card.
Knowledge cutoff is stated as March 2026 in the model card.
Multimodal input covers text, images, audio, and video.
Safety evaluation combined automated testing, manual red teaming for child safety, and a Frontier Safety Framework assessment.

Not disclosed: pre-training token count, compute budget, data mixture, the post-training recipe (supervised fine-tuning, reinforcement learning from verifiable rewards, distillation, or a combination), and any reward-model details. Some blogs describe the recipe with specifics. We could not trace those specifics to Google, so they are not repeated here.

What the behavior tells us about post-training

We can reason about the post-training signal without claiming its recipe. The model takes more turns, calls tools more often, and spends about 30% more output tokens per task. Those are the signatures of training that rewards task completion over brevity, typically on coding and agentic environments where a final pass or fail is the reward. That is an inference from observed behavior and Google’s description, not a disclosed fact.

The benchmark selection reinforces it: the launch materials lead with DeepSWE v1.1, Terminal-bench 2.1, Vals Finance Agent v2, and Harvey’s legal agent benchmark, all multi-step, tool-using evaluations. The weak spots (covered next) are also agentic, but of a different shape: broad, general-agent and computer-use tasks.

Capabilities and Benchmarks

Everything in this section is either from Google’s launch materials as reproduced by a named reviewer, or from Artificial Analysis. Where I could only verify a number through one secondary reviewer, I say so. Google’s launch post itself gives only some values in prose, such as 54.9% on HLE-Verified, over 70% on its internal vulnerability discovery benchmark, and 47.2% pass@1 on CWE-Bench.

The Google-reported comparison table

The table below reproduces the comparison figures published in Google’s launch table as transcribed by Vellum. Opus 5 refers to Anthropic’s Claude Opus 5 and Sol to OpenAI’s GPT-5.6 Sol, the comparison models in that table.

Benchmark	Gemini 3.8 Flash	Claude Opus 5	GPT-5.6 Sol	Note
DeepSWE v1.1 (coding)	73.7%	74.0%	72.7%	Within 0.3 points of Opus 5
Terminal-bench 2.1	89.4%	89.1%	88.8%	Highest in table
Vals Finance Agent v2	61.4%	58.6%	53.8%	Highest published
Harvey Legal Agent (all-pass)	10.0%	6.7%	2.5%	Low absolute rates
HLE-Verified	54.9%	54.4%	54.5%	Statistical near-tie
CharXiv (chart reasoning)	86.2%	83.7%	85.8%	Highest in table
LVBench (long video)	87.8%	75.4%	82.1%	Agentic mode
GDPVal-AA v2 (Elo)	1545	1824	1710	279 behind Opus 5
OSWorld-2.0 (computer use)	59.0%	75.4%	62.6%	Behind both
Terminal-bench 4.0 (general agent)	19.1%	51.8%	37.3%	Largest gap
Gray Swan prompt injection (attack success, lower is better)	5.5%	4.8%	27.0%	Second-best of three

The pattern is consistent. Gemini 3.8 Flash is at or near the top on coding, terminal, finance, chart, and video benchmarks. It trails badly on broad general-agent work and on knowledge-work Elo. A model that lands within 0.3 points of a far pricier competitor on DeepSWE but 32.7 points behind on Terminal-bench 4.0 is telling you that “agentic” is not one skill.

Independent measurement: Artificial Analysis

Artificial Analysis scored Gemini 3.8 Flash at high reasoning at 59 on its Intelligence Index, versus 56 for Gemini 3.7 Flash at high, a three-point gain concentrated in agentic and coding tasks. Their reported figures by reasoning level:

Reasoning level	Index score	Cost per task	Time per task
High	59	$0.58	about 2.5 min
Medium	57	$0.41	not reported here
Low	52	$0.24	about 0.8 min

Artificial Analysis says high reasoning generates about 48k output tokens per task, up roughly 30% from 3.7 Flash, at about 300 tokens per second. It called the model the cheapest it had measured at this intelligence level while noting the cost rose about 40% from 3.7 Flash’s $0.40.

What the caveats say

Vendor tables are selected. Google chose which benchmarks and which competitors to print. Absences (no 3.5 Flash column, no Gemini Pro column in the transcribed table) are informative.
Benchmark versions are new. DeepSWE v1.1, Terminal-bench 4.0, GDPVal-AA v2, and OSWorld-2.0 are recent versions; scores are not comparable to older versions of the same names.
Harness matters. The OSWorld figure carries a “batch tool setting” note; LVBench is “agentic mode.” Different scaffolds shift results by more than the gaps between models.
HLE-Verified is a near-tie. 54.9 versus 54.4 versus 54.5 is inside plausible run-to-run noise. Do not read a ranking into it.
Latency is a separate axis. One review citing Artificial Analysis reports about 13.3 seconds to first token against a 2.99-second median for comparable models. Reasoning models spend that time thinking; for an interactive UI it is a real constraint, whatever the throughput number says.

The Pricing Cliff: Cost Per Task, Not Cost Per Token

The per-token schedule is simple. Everything after it is not. This section builds the cost model that the launch pricing hides.

The published schedule

From Google’s pricing page, standard paid tier, per million tokens:

Item	Through Dec 31, 2026	From Jan 1, 2027
Input	$0.75	$1.50
Output (includes reasoning)	$3.75	$7.50
Cached input	$0.075	$0.15
Cache storage per hour	$0.50	$1.00
Batch input	$0.375	$0.75
Batch output	$1.875	$3.75

Every line doubles. Nothing is restructured, no tier is removed, and cached input keeps its 90% discount relative to fresh input. Gemini 3.7 Flash has identical pricing on the standard, batch, and flex tiers, so the cliff hits both models.

Decomposing the $0.58 task

Artificial Analysis reports $0.58 per Intelligence Index task at high reasoning and about 48k output tokens per task. Output tokens at $3.75 per million cost about $0.18 (48,000 x $3.75 / 1,000,000). That leaves roughly $0.40 of the measured cost attributable to input, meaning prompts, tool results, and repeated context across turns.

Two things to notice. This split is my derivation; Artificial Analysis does not publish it in the sources I could access, and its cache assumptions are unknown. And if the residual were entirely uncached input at $0.75 per million, it implies about 533,000 input tokens per task, which is plausible for a multi-turn agent re-reading its context. The lesson holds regardless of the exact split: on agentic work, input and context can outweigh visible output.

That is why the “output is 5x input price” intuition misleads. Output price is higher per token, but agent loops multiply input tokens by the number of turns. More turns, the behavior Google calls “works harder,” raises the input bill even when output grows only 30%.

The cliff, worked through

If token counts stay constant, cost doubles exactly at the boundary. Illustrative arithmetic using the published per-task costs:

Scenario (per task)	Today (intro)	From Jan 1, 2027
3.8 Flash, high	$0.58	about $1.16
3.8 Flash, medium	$0.41	about $0.82
3.8 Flash, low	$0.24	about $0.48
3.7 Flash, high	$0.40	about $0.80

At 100,000 tasks per month, illustrative: 3.8 Flash high costs about $58,000 now and $116,000 after the cliff. 3.7 Flash high goes from about $40,000 to $80,000. 3.8 Flash at low effort goes from about $24,000 to $48,000. The relative gap between 3.8 high and 3.7 high stays at about 1.45x before and after, because the cliff applies to both.

One analysis states that a $0.58 task becomes about $1.16 and is “nearly triple” the cost of the same workload on 3.7 Flash today. The arithmetic is right but the comparison mixes dates: it sets January 2027 prices for one model against today’s prices for the other. Like for like, both double.

Figure 3: Where the $0.58 goes and how the January 2027 cliff scales it. The input-versus-output split is derived and illustrative, not published.

The decision rule: cost per solved task

Cost per task is only half the metric. Builders should track cost per successful task: cost divided by success rate. Under this rule, 3.8 Flash at high effort beats 3.7 Flash at high effort only if its success rate is at least 1.45 times higher (0.58 / 0.40).

Illustrative example: suppose 3.7 Flash solves 60% of your tasks at $0.40. Cost per solved task is $0.40 / 0.60 = $0.67. For 3.8 Flash at high effort to match that, it must solve 0.58 / 0.67 = about 87% of them. At medium effort ($0.41), it needs only about 61.5%. These success rates are invented to show the mechanism; measure your own.

The independent index gain is three points, 56 to 59. A three-point index gain does not by itself justify a 45% cost premium. It may for workloads where 3.7 fails hard (Artificial Analysis-cited comparisons mention gains of roughly 12 points on banking tool-use and 8.4 points on software engineering evaluations, as reported by a third-party reviewer). For classification, extraction, and routing tasks, it will not.

Levers that change the math

Effort level. Low effort at 3.8 Flash scores 52 at $0.24; that beats the 3.7 Flash cost by 40%. Artificial Analysis notes it matches Gemini 3.6 Flash high at about 30% lower cost. Medium scores 57 at $0.41, close to 3.7 Flash’s price with a higher score than 3.7’s 56 at high. On these figures, medium is the cost-parity setting.
Context caching. Cached input is $0.075 per million against $0.75 fresh. Illustrative: if 70% of the roughly $0.40 input cost were a stable prefix that hits cache, input cost drops to about $0.15 and the task to about $0.33. Storage of $0.50 per million tokens per hour adds a carrying cost for long-lived caches; a 500k-token cache held for eight hours costs about $2.00 in storage at intro rates ($0.50 x 0.5 x 8), so it only pays off if reused often.
Batch tier. Batch rates are exactly half of standard ($0.375 input, $1.875 output). For non-interactive pipelines such as nightly code-review sweeps or bulk document analysis, batch turns the $0.58 task into roughly $0.29 with the same tokens (illustrative, assuming batch tokens match).
Output caps. With thinking_level replacing thinking_budget, you cannot cap reasoning tokens directly. Cap the total via max_output_tokens (limit 65,536), turn limits in your agent loop, and step budgets.

Budgeting for January

Treat January 2027 pricing as your planning baseline. Finance teams that budget on intro rates will see a 2x variance without any change in traffic. Three practical steps: instrument per-route token counts now, tag every agent run with effort level and model, and set alert thresholds on cost per solved task rather than on spend alone. For broader tactics, our inference cost optimization guide and the small-versus-large LLM agentic benchmark show how routing between model sizes changes blended cost.

Google has not said the introductory pricing will be extended, nor that it will not. The dates in the docs are the only commitment; plan on them.

Access and Deployment

Figure 4: Access routes. The general model is broadly available; the Cyber variant is reachable only through vetting.

Where to get it

Google’s launch post lists the surfaces. For developers: Google AI Studio, Android Studio, and Google Antigravity. For enterprises: Gemini Enterprise. For consumers: the Gemini app, AI Mode in Google Search, and Google Sheets. The API documentation adds that 3.8 Flash is now the default model for the Antigravity managed agent. The pricing page lists a free tier for the model, subject to its limits and to Google’s data-use terms for free usage; check those terms before sending anything sensitive.

Open weights and self-hosting

There are no open weights for Gemini 3.8 Flash. It is an API and product-surface model under Google’s terms of service; no license grants self-hosting, no VRAM figure applies, and quantization is not user-controllable. If your requirements include on-premises inference, air-gapped operation, or fixed-cost hosting, this is the wrong model class. That is a real trade-off against open-weight competitors, whose price schedules do not step on a date.

Rate, latency, and throughput

Artificial Analysis measured about 300 output tokens per second at high reasoning and about 2.5 minutes to complete an index task at high effort versus 0.8 minutes at low. Time to first token is long for a reasoning model, since thinking happens before the answer starts. If you build chat, budget for a streaming UX that shows progress, or run low effort for interactive paths and reserve high effort for background agents.

Migration notes

Migrating from 3.7 Flash is described in the API notes as changing the model string to gemini-3.8-flash. For code written against older Gemini conventions, expect to remove sampling parameters and thinking_budget, and to replace prefilled model turns with previous_interaction_id. Test tool-calling loops in staging: more iterations per task means more chances to hit your own timeouts and rate limits.

Gemini 3.8 Flash Cyber and the Fairwind Program

Google shipped a second model on the same day: Gemini 3.8 Flash Cyber. Google says it “ships with a more permissive set of mitigations for cybersecurity, and as such, is only available to trusted defenders who require a more comprehensive set of cyber capabilities.” Access is through the new Fairwind Program, offering “trusted government authorities, as well as critical infrastructure operators and software maintainers with prioritized access.”

You cannot buy it on the standard API. There is no public price for it in the sources I could verify, and eligibility criteria beyond those categories are not published in the launch post.

Reported results

Google’s launch post and its transcription by Vellum report:

CyberGym (vulnerability discovery): 86.2% pass@1, against 85.6% for GPT-5.5 Cyber and 83.8% for Mythos 5.
Internal 20-language vulnerability discovery benchmark: success rate above 70% (71.0% in the transcription).
CWE-Bench (patching): 47.2% pass@1, against 47.8% for the leading frontier model, so Cyber does not lead there.
Chrome Security team: 2.6 times more correct patches than “the best commercial models.”
Wiz: 7.5 to 9.7 percentage points higher recall on its internal penetration-testing benchmark at 2.3 to 5.2 times lower cost.
Google Cloud Vulnerability Research: found a critical foundational vulnerability in under two hours.

How to read it

These are vendor-reported, mostly on internal or partner benchmarks that outsiders cannot rerun. The Chrome and Wiz figures come from named partners but not from published methodology. The honest reading: strong signals of capability, without independent replication. Note also that Cyber is not uniformly best: on CWE-Bench it trails a competitor by 0.6 points.

The gating model is the more interesting design decision. The same base model ships in two policy configurations: a general one with strict cyber refusals, and a restricted one with looser mitigations for vetted users. Safety here is enforced by access control and identity, not only by weights. That has consequences for defenders: capability that sits behind vetting is not available to smaller security teams unless they qualify. And for red-team planning, assume adversaries will eventually reach comparable capability elsewhere. Our note on agentic AI security and prompt injection covers the defensive side of tool-using models.

On prompt injection, the launch post cites AI security firm Gray Swan for a “significant leap” in robustness. In the transcribed comparison table, 3.8 Flash’s attack success rate is 5.5% versus 4.8% for Opus 5 and 27.0% for Sol (lower is better). Even 5.5% means roughly one in eighteen crafted attacks succeeds in that test; treat untrusted content as hostile regardless.

Limitations, Safety, and Failure Modes

Verbosity is the failure mode that costs money

The clearest weakness is not wrong answers; it is token appetite. One review citing Artificial Analysis says the model consumed about 120 million output tokens to run the index suite against a median of 71 million for comparable models, roughly 69% above typical. Total evaluation cost was reported at $825.83. A model that thinks more helps on hard tasks and wastes money on easy ones, and the minimal thinking level that would help is unsupported.

Latency and interactive use

Reported time to first token is long, about 13.3 seconds against a 2.99-second median in one summary of Artificial Analysis data. That is disqualifying for keystroke-level assistance and borderline for chat. Use low effort or 3.7 Flash on interactive paths.

Broad agent and computer-use gaps

Terminal-bench 4.0 at 19.1% versus 51.8% for Opus 5, and OSWorld-2.0 at 59.0% versus 75.4%, show that top marks on coding and terminal benchmarks do not transfer to open-ended agents or GUI control. GDPVal-AA v2 at 1545 Elo trails Opus 5’s 1824. If your agent drives a desktop or handles ambiguous multi-department workflows, test against a larger model before committing.

Long-context degradation

A 1M-token window does not guarantee 1M tokens of reliable recall. Google has not published a needle-in-a-haystack or multi-hop retrieval curve for 3.8 Flash that I could verify. Assume accuracy drops as relevant evidence gets diluted, and evaluate on your own documents. Retrieval or summarization that keeps prompts at tens of thousands of tokens remains cheaper and usually more accurate.

Safety and misuse

Google reports safety and tone similar to 3.7 Flash, low unjustified refusals, and no Frontier Safety Framework threshold concerns. Those are self-reported evaluations. Standard cautions apply: hallucination remains possible on obscure facts, the March 2026 knowledge cutoff leaves gaps, and prompt injection is reduced but not eliminated. For the Cyber variant, the safety control is vetting, so a compromised or careless Fairwind user is a supply-chain risk for that capability.

Vendor and pricing risk

The price schedule is time-limited by design, the model line turns over in weeks, and the API surface changed (thinking levels, server-side state). Lock-in risk is real: previous_interaction_id state lives with the provider. Keep an abstraction layer and a fallback model.

How Gemini 3.8 Flash Compares

Peers priced from the Vellum comparison: Claude Opus 5 at $5/$25 per million tokens, GPT-5.6 Sol at $4/$20, and Claude Sonnet 5 at $2/$10. Those figures come from one reviewer; verify current prices before relying on them. Scores are from the vendor table above.

Use case	Gemini 3.8 Flash	Gemini 3.7 Flash	Claude Opus 5	GPT-5.6 Sol
Autonomous coding on long tasks	Strong: 73.7% DeepSWE at lowest list price of the group	Cheaper per task ($0.40), lower index	74.0%, highest list price	72.7%
High-volume extraction and classification	Overkill at high effort; use low	Best default	Overkill	Overkill
General agents and computer use	Weak: 19.1% and 59.0%	Untested here, expect weaker	Strongest in table	Middle
Long-video and chart analysis	Best in table on CharXiv and LVBench	Not compared	Behind	Behind on both
Interactive chat	Slow first token	Better choice	Not measured here	Not measured here
Regulated finance or legal agents	Best published Vals Finance Agent score	Not compared	Behind on the finance score	Behind on both

The consistent message: 3.8 Flash wins on price-performance for structured, tool-heavy tasks and loses on breadth. For the OpenAI comparison see our GPT-5.6 Sol explainer; for agent benchmark background, see AI agent benchmarks such as SWE-bench, GAIA, and tau-bench.

Trade-offs, Gotchas, and What Goes Wrong

Optimizing on the wrong number. Choosing the model on $/M tokens produces a 40% surprise at high effort. Optimize on cost per solved task.
Ignoring the cliff. A prototype that looks affordable in October can double in January. Model both prices in your business case now.
Turn explosion. More tool calls per task can hit your rate limits, tool timeouts, or downstream API costs (search, code sandboxes) that do not shrink when tokens get cheaper.
Effort creep. Teams default to high because it scores best, then never revisit. Add an automated evaluation that checks whether medium gives the same pass rate on your tasks.
Benchmark transplant. DeepSWE success on curated repositories is not your monorepo’s success. Run 50 to 100 of your own tasks before migrating.
Cache misconfiguration. Cache storage is billed hourly; an unused cache is a leak. Keep prefixes stable (system prompt, tool schemas first) so hits are frequent.
Assuming Cyber access. Do not architect a product around Gemini 3.8 Flash Cyber unless your organization qualifies for Fairwind.

Practical Recommendations

Start from your workload, not the leaderboard. If your work is high-volume, simple, and latency-sensitive, stay on 3.7 Flash or use 3.8 Flash at low effort; Google itself says 3.7 Flash remains fully supported for efficiency-first workloads. If your work is long-horizon coding or tool-heavy analysis where failures are expensive, run 3.8 Flash at medium first, and promote to high only where medium’s failures are measurable.

Do the evaluation in dollars. Run the same 50 to 100 real tasks across 3.7 Flash, 3.8 Flash at each level, and one larger model. Record success, tokens in, tokens out, turns, wall-clock time, and cost at both today’s and January 2027 prices. Pick on cost per solved task.

Then instrument. Tag each agent run, cap turns, and alert on drift. Use batch for anything that can wait: it halves the price on both sides of the cliff.

Checklist:

Budget on January 2027 rates, not intro rates.
Benchmark 3.7 Flash, 3.8 Flash low, medium, and high on your own tasks.
Compute cost per solved task, and require success-rate gains of 1.45x to justify high over 3.7 high.
Use context caching for stable prefixes; use batch for offline work.
Keep low effort on interactive paths; measure time to first token.
Cap agent turns and total output tokens.
Keep a fallback model and a provider abstraction.
Treat all tool-returned content as untrusted.

Frequently Asked Questions

What is Gemini 3.8 Flash and when did it launch?

Gemini 3.8 Flash is Google’s Flash-tier model, announced on September 2, 2026 alongside a gated Cyber variant. Its API model ID is gemini-3.8-flash. It accepts text, images, audio, and video, outputs text, and has a 1,048,576-token input window with up to 65,536 output tokens. It is aimed at agentic coding and tool-heavy workflows, and is generally available through the Gemini API and Google’s products.

How much does Gemini 3.8 Flash cost?

Through December 31, 2026, it costs $0.75 per million input tokens and $3.75 per million output tokens, including reasoning tokens. From January 1, 2027, prices double to $1.50 and $7.50. Cached input is $0.075 (then $0.15), and the batch tier is half the standard rate. Real spend depends on tokens per task: Artificial Analysis measured $0.58 per task at high reasoning.

Is Gemini 3.8 Flash cheaper than 3.7 Flash?

Per token, no: the prices are identical. Per task, it can be more expensive. At high reasoning, Artificial Analysis measured $0.58 versus about $0.40 for 3.7 Flash, roughly 40% more, because 3.8 generates about 30% more output tokens and takes more agent turns. At medium reasoning the cost is about $0.41, close to parity. Both models double in price in January 2027.

Can I use Gemini 3.8 Flash Cyber?

Only if you qualify. Google offers it through the Fairwind Program, aimed at trusted government authorities, critical infrastructure operators, and software maintainers, with more permissive cybersecurity mitigations than the general model. It is not on the standard public API, and Google has not published pricing or full eligibility criteria in its launch post. Everyone else uses the general Gemini 3.8 Flash, which keeps stricter cyber safeguards.

How does Gemini 3.8 Flash compare with Claude Opus 5 and GPT-5.6 Sol?

In Google’s published table, it scores 73.7% on DeepSWE (Opus 5: 74.0%, Sol: 72.7%), 89.4% on Terminal-bench 2.1, and 54.9% on HLE-Verified, a near-tie. It trails badly on Terminal-bench 4.0 (19.1% vs 51.8% for Opus 5) and OSWorld-2.0. It is much cheaper per token, but higher token use narrows the cost gap. These are vendor-reported figures; validate on your workload.

Does Gemini 3.8 Flash have open weights or self-hosting?

No. It is available only as a hosted Google service through the Gemini API, Google AI Studio, Gemini Enterprise, and Google consumer products. There is no downloadable checkpoint, no license for on-premises deployment, and no published parameter count. Teams that need self-hosting, air-gapped operation, or a fixed price schedule should evaluate open-weight alternatives instead and compare their total cost of ownership.