by Google · Gemini 3.8 family · best for long-horizon coding and agent workhorse
Gemini 3.8 Flash is Google's newest Flash model, released GA on 2026-09-02 as the third Flash release in six weeks (after 3.6 Flash on 2026-07-21 and 3.7 Flash on 2026-08-13). Google calls it its "most intelligent Flash model", built for long-horizon software engineering, autonomous agents and enterprise knowledge work, and it is now the default model behind the Antigravity agent and the Antigravity SDK. On Google's own table it matches or beats GPT-5.6 Terra and Claude Sonnet 5 on most rows and sits close to Claude Opus 5 on coding and reasoning, at an introductory $0.75 in / $3.75 out per 1M tokens. That price runs through 2026-12-31; from 2027-01-01 it doubles to $1.50 / $7.50. For a buyer: this is the strongest price-to-capability model on Google's stack today, with two catches: it deliberately spends more tokens on hard tasks, and the sticker price doubles in January.
temperature, top_p and top_k sampling parameters.low, medium (default) and high; minimal returns an error.| Benchmark | Score | Source |
|---|---|---|
| LMArena Elo | 1493 | lmarena.ai 2026-09-13T00:00:00.000Z |
| Terminal-Bench | 89.4% | deepmind.google 2026-09-02T00:00:00.000Z |
| LMArena Coding Elo | 1535 | lmarena.ai 2026-09-13T00:00:00.000Z |
Six personas, six verdicts — the same panel that reviews every product on TopReviewed.
“A GA Flash that trades blows with Opus 5 on coding for a fraction of the cost. Standardize on it, but budget for the January 2027 price doubling.”
For a CTO, 3.8 Flash is the easiest recommendation on Google's stack right now. It is GA with no shutdown date, it is the default model for Google's own managed agents, and Google's table puts it level with or ahead of GPT-5.6 Terra and Claude Sonnet 5 on most rows. Governance is the full Agent Platform set: regional endpoints, CMEK, VPC-SC and Access Transparency. The risks are cadence and price. Google shipped three Flash models in six weeks, and the rate doubles on 2027-01-01. Lock-in is to Google Cloud and the Interactions API, not to this specific model.
“Google's Flash tier now contests the mid-market Terra and Sonnet 5 own, at a lower price, and it wins most of the rows Google chose to publish.”
3.8 Flash repositions the Flash tier from cheap helper to near-frontier workhorse. On Google's table it beats GPT-5.6 Terra on 12 of 13 benchmarks (all but Terminal-Bench 4.0) and Claude Sonnet 5 on all but GDPval-AA, and it leads specialist agent benchmarks in finance, legal and biology. Its moat is distribution: Search grounding, the Gemini app, AI Mode, Sheets and Antigravity. The competitive threats are Grok 4.7, which xAI reports ahead on Terminal-Bench 4.0 and Harvey at $2 / $6, and DeepSeek V4.1-Flash, which undercuts it heavily with open weights. The timing is strong, but the introductory price makes the positioning temporary until Google confirms 2027 pricing.
“At $0.75 in and $3.75 out it is a bargain through December. Model it at $1.50 and $7.50, and assume it uses more tokens per task than 3.7.”
On a 10M-input, 2M-output daily load, 3.8 Flash costs $15.00 at the introductory rate against $44.00 for GPT-5.6 Terra, $40.00 for Claude Sonnet 5 and $20.00 for Claude Haiku 4.5. At the standard rate the same load is $30.00, still below Terra, Sonnet 5 and Grok 4.7 ($32.00), but far above DeepSeek V4.1-Flash at $5.40. Batch and Flex halve the rate and caching cuts input by 90%. The unknown is token volume: Google says 3.8 Flash can spend more tokens by design. Value is 9.4 now and about 8.7 at standard pricing.
“Tool loops are sturdier and it checks its own work, but migrating means rewriting configs: no temperature, no thinking budget, no minimal effort.”
For builders, 3.8 Flash offers the full Gemini toolset in one GA model: function calling, structured output, code execution, computer use, Search and Maps grounding, URL context and file search. It is also the default model for the Antigravity agent. Google says it cuts failed loops by planning in smaller steps and verifying as it goes. Migration takes real work. Sampling parameters are gone, thinking_budget becomes thinking_level, minimal errors, and multi-turn state moves to previous_interaction_id. Some developers will also hit Malformed_Function_Call on pre-tool text. The docs are thorough and include a coding-agent migration skill.
“Sits in LMArena's top ten next to Opus 5, and answers feel thorough, though the default medium effort can make simple questions slower than 3.7 Flash.”
3.8 Flash is available to Google AI Pro ($19.99/mo) and Ultra subscribers in the Gemini app, AI Mode in Search and Gemini in Sheets. Free users get 3.6 Flash in the app. Its preliminary LMArena text rating of 1493 ties Claude Opus 5, which suggests people like its answers. Multimodal input, charts, long video and documents are strengths. The trade-off is effort: it reasons more before answering, so quick questions can feel slower than on 3.7 Flash. Google reports low unjustified refusals, only slightly up on 3.7 Flash.
“Every headline number is Google's own, the one independent rating is preliminary, and the price that makes it look cheap expires on 2026-12-31.”
Treat the launch table as a vendor document. Gemini scores are mostly self-computed, competitor figures are mostly self-reported, and Google chose the benchmarks. Where it loses, it loses badly: Terminal-Bench 4.0 is 19.1% against Opus 5's 51.8%, and OSWorld-2.0 is 59.0% against 75.4%. xAI reports Grok 4.7 far ahead on Terminal-Bench 4.0 and Harvey. The HLE-Verified lead is under half a point. The only independent signal is a preliminary LMArena rating with a rank spread of 1 to 24, and there is no independent speed or token-use data. The model spends more tokens by design, and the price doubles in January. Expect strong results, and verify them on your own workloads.
minimal effort level; Google warns of occasional slowness or timeouts.The full research notes behind this review — verified against primary sources.
Google publishes no parameter count, expert count, layer count or attention design for 3.8 Flash, so those fields stay null. The model card traces the lineage: 3.8 Flash is based on 3.7 Flash, which is based on 3.6 Flash, then 3.5 Flash and Gemini 3 Flash, whose card points to Gemini 3 Pro, described as a sparse mixture-of-experts transformer with native multimodal input. Hence architecture_type: moe. Disclosed and verifiable: a 1,048,576-token input window, 65,536 output tokens, a March 2026 knowledge cutoff with some domains stopping at January 2025, and text, image, audio and video input. Google says the gains over 3.7 Flash came from long-running agentic loops that evaluate and refine the model, plus heavy training on cybersecurity tasks. 3.8 Flash shares that foundation with Gemini 3.8 Flash Cyber, a restricted variant for vetted defenders.
Agentic work is the headline (cap_agentic 9.5, cap_function_calling 9.5). On Google's card 3.8 Flash leads every compared model on Vals Finance Agent v2 (61.4%), Harvey's Legal Agent Benchmark (10.0% all-pass) and Terminal-Bench 2.1 (89.4%), and it is the default model for Google's managed Antigravity agent. It loses clearly on longer, harder agent runs: Terminal-Bench 4.0 is 19.1% against Opus 5's 51.8%, and OSWorld-2.0 computer use is 59.0% against Opus 5's 75.4%. Coding (9.0) is close to the frontier: DeepSWE v1.1 is 73.7% against Opus 5's 74.0% and GPT-5.6 Sol's 72.7%. Reasoning (9.0) rests on HLE-Verified at 54.9%, the top score in Google's table, though only 0.4 to 0.5 points ahead of Opus 5 and Sol. Math (8.5) has no dedicated benchmark; the score leans on HLE-Verified's STEM share and the high effort level. Vision and video (9.5) lead the table: CharXiv Reasoning 86.2% and LVBench 87.8% agentic. Document understanding (9.0) is solid but not first: GDP.PDF is 35.0% against Sol's 40.0%. Long context (9.0) is inferred from lineage: 3.7 Flash scored 97.0% on 128K MRCR v2, while Google has published no long-context figure for 3.8 Flash itself. Real-time data (9.0) comes from Google Search and Google Maps grounding plus URL context. Creative writing (8.0) has no benchmark; its preliminary top-10 LMArena text rating is the only signal. Instruction following (8.5) and multilingual (8.5) carry over from the 3.5 Flash baseline. Safety calibration (8.0) reflects low unjustified refusals, a slight multilingual safety regression and improved prompt-injection robustness.
| Benchmark | Score | vs Predecessor | vs Top Competitor | Source |
|---|---|---|---|---|
| DeepSWE v1.1 (long-horizon SWE) | 73.7% | 3.7 Flash 65.3%; 3.5 Flash 37% | Opus 5 74.0%, GPT-5.6 Sol 72.7%, Terra 69.6%, Sonnet 5 53.8% | card |
| Terminal-Bench 2.1 | 89.4% | 3.7 Flash 85.8%; 3.5 Flash 76.2% | Highest in table: Opus 5 89.1%, Sol 88.8%, Terra 87.4% | card |
| Terminal-Bench 4.0 | 19.1% | 3.7 Flash 11.2% | Far behind Opus 5 51.8%, Sol 37.3%; behind Terra 23.6% | card |
| GDPval-AA v2 (Elo) | 1545 | 3.7 Flash 1482; 3.5 Flash 1349 | Behind Opus 5 1824, Sol 1710, Sonnet 5 1584; ahead of Terra 1528 | card |
| Vals Finance Agent v2 | 61.4% | 3.7 Flash 59.0% | Leads table; Opus 5 58.6%, Terra 54.4% | card |
| Harvey Legal Agent Benchmark (all-pass) | 10.0% | 3.7 Flash 8.8% | Leads Google's table (Opus 5 6.7%, Terra 0.8%) | card |
| GDP.PDF (all-pass) | 35.0% | 3.7 Flash 34.0% | Behind Sol 40.0%, Opus 5 37.0%; ahead of Terra 29.0% | card |
| CharXiv Reasoning (no tools) | 86.2% | 3.7 Flash 84.5%; 3.5 Flash 84.2% | Leads table; Terra 85.9%, Sol 85.8% | card |
| LVBench (long video) | 87.8% agentic / 87.1% static | 3.7 Flash 85.4% | Leads table; Sol 82.1%, Terra 78.9% | card |
| HLE-Verified | 54.9% | 3.7 Flash 53.6% | Leads narrowly: Sol 54.5%, Opus 5 54.4%, Terra 51.1% | card |
| OSWorld-2.0 (partial score) | 59.0% | 3.7 Flash 50.6% | Behind Opus 5 75.4%, Sol 62.6%; ahead of Terra 50.2% | card |
| BioMysteryBench (human-solvable / human-difficult) | 88.8% / 56.5% | 3.7 Flash 87.1% / 43.5% | Opus 5 90.1% on solvable; leads on difficult | card |
| LABBench2 | 86.2% | 3.7 Flash 82.1% | Leads table; Opus 5 84.2% | card |
| LMArena Text Elo (thinking high) | 1493 ±9, preliminary | 3.7 Flash 1490; 3.5 Flash 1478 | Ties Opus 5 high 1493; ahead of Terra xhigh 1466, Sonnet 5 high 1461 | Arena |
Every row except LMArena is vendor-reported: Google's model card and evaluation methodology, dated September 2026. Per that methodology, most Gemini scores are self-computed. Competitor figures are mostly the rivals' own published numbers. GDPval-AA v2 comes from the Artificial Analysis leaderboard, and the Vals Finance and Harvey figures come from Vals.AI. HLE-Verified is a 1,811-question cleaned subset of Humanity's Last Exam, not the original set, so it stays out of the hle field. The LMArena rating is independent but preliminary: 5,076 votes, rank 9 of 402 models at the 2026-09-13 vote cutoff, with a rank spread of 1 to 24. Google has not published GPQA Diamond, SWE-bench Verified or Pro, MMMU or MRCR figures for 3.8 Flash, and no Artificial Analysis Index figure could be read, so those fields are null.
No independent throughput or time-to-first-token measurement for 3.8 Flash could be read, so both fields are null. Google says it runs "at the same speed" as 3.7 Flash, but the model is designed to take more reasoning steps and tool calls on hard tasks, so wall-clock time per task can be longer than per-token speed suggests. Latency controls are effort levels (low for real-time chat and fast drafts, medium by default, high for deep reasoning) and consumption tiers: Priority inference for latency-sensitive traffic, Flex and Batch for cheaper asynchronous work. There is no minimal effort level, so every request pays some thinking cost. Google's model card also warns of "occasional slowness or timeout issues".
| Surface | Cost | Notes |
|---|---|---|
| API input | $0.75 / 1M tok | Introductory through 2026-12-31; $1.50 from 2027-01-01 |
| API output | $3.75 / 1M tok | Includes thinking tokens; $7.50 from 2027-01-01 |
| Cached input | $0.075 / 1M tok | $0.15 from 2027-01-01; storage $0.50 per 1M tok per hour ($1.00 from 2027-01-01) |
| Batch / Flex (in/out) | $0.375 / $1.875 | $0.75 / $3.75 from 2027-01-01 |
| Priority (in/out) | $1.35 / $6.75 | $2.70 / $13.50 from 2027-01-01 |
| Image, audio, video input | Same per-token input price | No separate audio rate on the pricing page |
| Search and Maps grounding | 5,000 requests per month free (shared across Gemini 3 models), then $14 per 1,000 | Paid tier only |
| Direct UI | Google AI Pro $19.99/mo; AI Ultra from $99.99/mo | 3.8 Flash in the Gemini app for Pro and Ultra; the Free plan gets 3.6 Flash |
| Free tier | Input and output free of charge in the Gemini API | Rate-capped; free-tier prompts are used to improve Google products |
| Rate limits | Not published on the model or pricing page |
Worked cost for a steady workload of 10M input and 2M output tokens a day, no caching, assuming equal token counts per task:
| Model | $ in / $ out per 1M | Daily cost |
|---|---|---|
| Gemini 3.8 Flash, introductory (to 2026-12-31) | 0.75 / 3.75 | $15.00 |
| Gemini 3.8 Flash, standard (from 2027-01-01) | 1.50 / 7.50 | $30.00 |
| Gemini 3.5 Flash | 1.50 / 9.00 | $33.00 |
| OpenAI GPT-5.6 Terra | 2.00 / 12.00 | $44.00 |
| Anthropic Claude Sonnet 5 | 2.00 / 10.00 | $40.00 |
| Anthropic Claude Haiku 4.5 | 1.00 / 5.00 | $20.00 |
| xAI Grok 4.7 | 2.00 / 6.00 | $32.00 |
| DeepSeek V4.1-Flash (peak) | 0.30 / 1.20 | $5.40 |
Value score is 9.4 at the introductory price and about 8.7 at the standard price. The standard price still undercuts Terra and Sonnet 5 at comparable or better benchmark results and roughly matches Grok 4.7 on this mix, though its $7.50 output rate is above Grok's $6.00, and it stays well above DeepSeek V4.1-Flash. Because 3.8 Flash spends more output tokens by design, real per-task cost runs above this table unless you lower the effort level.
Proprietary and closed-weights; no self-hosting. Access routes: the Gemini API through Google AI Studio (gemini-3.8-flash, Interactions and GenerateContent APIs), Android Studio, Google Antigravity (default model for the Antigravity agent and SDK) and Stitch. Enterprises use Gemini Enterprise and Gemini Enterprise Agent Platform (formerly Vertex AI). On Agent Platform, 3.8 Flash is GA with a global endpoint and US and EU multi-region endpoints. It supports Provisioned Throughput, batch inference and Standard, Flex and Priority PayGo, with data residency, customer-managed encryption keys, VPC Service Controls and Access Transparency on online prediction, batch and context caching. Agent Platform lists tuning as not supported, and its page marks agentic video understanding as not supported, although the Gemini API supports it. Since 2026-09-16, 3.8 Flash is also the default model in Google's CodeMender CLI. OpenRouter resells it as google/gemini-3.8-flash at Google's list price. Migration from earlier Gemini models means more than a model-ID swap. You drop temperature, top_p and top_k, replace thinking_budget with thinking_level (no minimal) and remove candidate_count. You also move multi-turn state to previous_interaction_id and check function calls for the documented Malformed_Function_Call pre-tool-text issue.
Covered by Google's Frontier Safety Framework. The 3.8 Flash card relies on 3.7 Flash's April 2026 assessment, which found no tracked or critical capability level reached; 3.7 Flash did reach the alert threshold for the cyber and CBRN uplift CCLs. Google says 3.8 Flash has no material new capability in those domains. It ships with CBRN and cyber-offense safeguards; the more permissive Cyber variant is limited to vetted defenders through the Fairwind Program. Google's automated safety results against 3.7 Flash: text safety 0.4 points better, image safety unchanged, multilingual safety 5.4 points worse, unjustified refusals 1.1 points higher. Google calls refusals low overall. Paid Gemini API and Agent Platform traffic is not used to improve Google products; free-tier traffic is. Gemini Enterprise Agent Platform is in scope for Google Cloud's SOC 2, ISO 27001 and HIPAA programs, and Google Cloud's GDPR commitments apply. The model card lists hallucination as a known limitation, so ground factual workloads with Search or retrieval.
Google GenAI SDKs for Python, JavaScript/TypeScript, Go and Java, plus REST. The model works through the Interactions and GenerateContent APIs and with LangChain, LlamaIndex, Vercel AI SDK, Genkit, Google ADK and Model Context Protocol tooling. It powers the Gemini app for Google AI Pro and Ultra subscribers, AI Mode in Google Search, Gemini in Google Sheets, Gemini Enterprise, the Antigravity agent and SDK (as default), Stitch and the CodeMender CLI (as default since 2026-09-16), and OpenRouter resells it. Distribution across Google's consumer and developer surfaces makes it mainstream from launch.
Google's pricing page lists $0.75 / $3.75 through 2026-12-31 and $1.50 / $7.50 from 2027-01-01. Cached, batch, Flex and Priority rates double on the same date. Google has not said whether it will extend the introductory rate, so forecast 2027 at the standard price.
Both cost the same. 3.8 Flash scores higher on every row of Google's card but spends more tokens on hard tasks. Google recommends 3.7 Flash, which stays fully supported, when compute efficiency matters more than peak quality.
On Google's own benchmarks it beats both on most rows, and LMArena's preliminary rating agrees, putting it about 30 points ahead. Terra wins on Terminal-Bench 4.0 and Sonnet 5 wins on GDPval-AA. Test on your own tasks, because nearly all the figures are vendor-reported.
Update the model ID, then remove temperature, top_p, top_k and candidate_count and replace thinking_budget with thinking_level. Google publishes a migration checklist and a coding-agent skill that automates most of it.
No. The lowest level is low, and minimal returns an error. Use low for latency-critical chat and drafts, and keep medium (the default) for code and agents.
Paid Gemini API and Agent Platform traffic is not used to improve Google's products; free-tier traffic is. Agent Platform adds data residency (global, US or EU), CMEK, VPC-SC and Access Transparency.
The window is 1,048,576 tokens, but Google has published no long-context retrieval score for 3.8 Flash. Its predecessors scored 97.0% (3.7 Flash) at 128K and 54.0% (3.6 Flash) at 1M on MRCR v2. Treat recall near the full window as unproven and chunk precision-critical retrieval.
Grok 4.7 ($2 / $6, 500K context) is xAI's flagship. xAI reports it ahead of 3.8 Flash on Terminal-Bench 4.0 (38.0% vs 19.1%) and Harvey, and behind it on DeepSWE (71.0% vs 73.7%). Claude Haiku 4.5 ($1 / $5, 200K context) is cheaper than 3.8 Flash's standard rate but a tier below in capability.
OpenAI's mini-tier model at $2 / $12 with a 1.05M context and 128K output. Google's table has 3.8 Flash ahead on DeepSWE (73.7% vs 69.6%), OSWorld-2.0 (59.0% vs 50.2%) and HLE-Verified (54.9% vs 51.1%), and LMArena rates it about 27 points higher. Terra wins on Terminal-Bench 4.0 (23.6% vs 19.1%) and on maximum output length. 3.8 Flash is cheaper even at its 2027 standard price.
$2 / $10, a price Anthropic made permanent after its own introductory period. Sonnet 5 wins GDPval-AA knowledge work (1584 vs 1545) and allows 128K output. 3.8 Flash wins DeepSWE (73.7% vs 53.8%), Terminal-Bench 2.1, CharXiv, LVBench and HLE-Verified, where Google notes many Sonnet 5 answers were blocked by content filters, and it costs less at both of its rates.
MIT open weights, 1M context and image input at $0.30 / $1.20 peak ($0.15 / $0.60 off-peak). That is roughly a third of 3.8 Flash's introductory price and a fifth or less of its standard price, and it can be self-hosted. 3.8 Flash wins on audio and video input, Search grounding, enterprise controls and published agentic results. The two vendors publish no shared head-to-head benchmark.
Primary references used to verify this review.
Does not train on API inputs by default
Last verified 2026-09-21