gemini-3.8-flash
Meant to excel at
-
Google's most intelligent Flash workhorse, improved over 3.7 Flash in software engineering, agentic tasks and multi-step domain reasoning.
Vendor's own site blog.google -
Targets long-horizon coding and autonomous agents; DeepSWE v1.1 73.7% is near Opus 5 (74.0%) at far lower cost per task.
Vendor claim, relayed vellum.ai -
Leads or ties on finance and legal agent benchmarks (Vals Finance Agent v2 61.4%, Harvey legal all-pass 10.0%).
Vendor claim, relayed vellum.ai -
Very fast output (about 302 to 305 tokens/s, third fastest tracked) and strong prompt-injection resistance (5.5% attack success).
Independent evaluator eesel.ai -
Works best inside a multi-turn agentic harness like Antigravity; scoped, cost-sensitive coding is the sweet spot.
Independent evaluator mindstudio.ai -
Caught and fixed all three planted bugs in each of three trials and flagged extra real issues.
Practitioner report developer.puter.com -
Rank 3 on the Terminal-Bench 2.1 aggregator at 89.4%, ahead of GPT-5.6 Sol and Kimi K3 (harness varies per entry).
Independent evaluator anotherwrapper.com
Watch out for
-
Weak on hard agentic terminal and computer-use work: Terminal-Bench 4.0 19.1% (Opus 5 51.8%), OSWorld-2.0 59.0% (75.4%).
Independent evaluator vellum.ai -
Verbose, with about 13s time to first token; roughly 30% more output tokens per task than 3.7 Flash.
Independent evaluator eesel.ai -
Introductory price of $0.75/$3.75 doubles on 1 January 2027, and extra thinking tokens bill at the output rate.
Vendor's own site blog.google -
Output quality drops in plain chat versus an agentic harness, and visual design instincts trail Claude models.
Independent evaluator mindstudio.ai -
Scored 44 vs 97 for 3.7 Flash on one rate-limiter task after rejecting a zero refill rate; single task and arguable prompt.
Practitioner report getneotiler.com
Benchmarks
Agentic coding percentile 0.64
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| deepswe v1.1 | 73.7 | 0.60 | Vendor's own site deepmind.google | 2026-09-02 |
| terminal-bench 2.1 | 89.4 | 0.91 | Vendor's own site deepmind.google | 2026-09-02 |
| terminal-bench 2.1 (aggregator, anotherwrapper.com) | 89.4 | 0.92 | Independent evaluator anotherwrapper.com | 2026-10-19 |
| terminal-bench 4.0 | 19.1 | 0.13 | Vendor's own site deepmind.google | 2026-09-02 |
Overall percentile 0.65
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 59 | 0.96 | Independent evaluator theregister.com | 2026-09-02 |
| gdpval-aa v2 | 1545 | 0.33 | Vendor's own site deepmind.google | 2026-09-02 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- -
Pro-tier user says Flash lost context across files, looped, hallucinated deprecated APIs and dropped code during refactors.
Compared with gemini-3.1-pro | 2026-10-14
Practitioner report discuss.ai.google.dev - +/-
Flash is faster, but a simple function move silently dropped a line of code and cost debugging time.
Compared with gemini-3.1-pro | 2026-10-14
Practitioner report discuss.ai.google.dev - +
A Google DeepMind engineer says 3.8 Flash takes smaller steps and verifies its work more often, which burns more tokens but helps long tasks (vendor-side framing).
Task: agentic-long-running | Compared with gemini-3.7-flash | 2026-09-08
Practitioner report eesel.ai - -
An X user called it unusable for SWE work and said it cost more than another model on simple coding tasks; another tester noted it cannot test its own work in the Antigravity browser.
Task: feature | 2026-09-08
Practitioner report eesel.ai - +
Simon Willison reported an HTML visualization generated in 13 seconds for about 1.8 cents and called the speed and frontend quality exciting.
Task: feature | 2026-09-08
Practitioner report eesel.ai - +/-
HN commenters say it is fast and holds quality under weekday load, but verbose (about 11k more tokens per task than 3.7 Flash), requires constant approvals in the CLI, and is fine for scaffolding but not long agentic tasks.
Task: agentic-long-running | Compared with gemini-3.7-flash | 2026-09-08
Practitioner report eesel.ai - -
On a thread-safe token-bucket rate limiter with pytest tests, 3.8 Flash wrote the most polished code but scored 44/100 on an external harness (3.7 Flash 97, 3.1 Pro 100) because it rejected a zero refill rate. Single task; the author's justification for that requirement is questionable.
Task: feature | Compared with gemini-3.1-pro | 2026-09-04
Practitioner report getneotiler.com - +/-
The same detailed prompt gave a thinner result in the Gemini app than in Antigravity, whose implement-verify-revise loop produced one of the best outputs, a live ISS tracker.
2026-09-03
Practitioner report mindstudio.ai - +/-
A data-center slide deck had correct brand colors and structure, but the author suggested a Claude model for design-heavy decks; product landing pages were uneven.
Compared with claude-fable-5.1 | 2026-09-03
Practitioner report mindstudio.ai - +
Given a short Python function with a SQL injection, a mutable default argument and a divide-by-zero, 3.8 Flash found and correctly fixed all three in all three runs, adding extra valid findings; latency ranged 18.6-41.1s. No other model was tested.
Task: bug-fix | 2026-09-03
Practitioner report developer.puter.com - -
One X user called it unusable for software engineering and said it cost more than Fable on simple coding tasks.
Compared with claude-fable-5 | 2026-09-02
Practitioner report eesel.ai - +/-
Praised its price, speed and reasoning, but noted it cannot test its own work or view the browser in Antigravity.
2026-09-02
Practitioner report eesel.ai - +
A commenter used the prior 3.7 Flash to audit Opus output and said it finds subtle defects even when unit tests pass, and runs far faster in Antigravity CLI.
Task: code-review | Compared with claude-opus-5 | 2026-08-14
Practitioner report eesel.ai - +
Glean's AI lead reportedly saw more than three times as many completed tasks than with 3.7 Flash on long document workflows.
Compared with gemini-3.7-flash undated
Practitioner report eesel.ai - +/-
Gateway tests over 150 cases found parity with 3.7 Flash on function calling (including parallel), structured JSON output and thinking tiers.
Compared with gemini-3.7-flash undated
Practitioner report docs.apiyi.com