Google DeepMind released Gemini 3.8 Flash on September 2. According to the official model card, it builds on 3.7 Flash, with improvements concentrated in software engineering and agentic knowledge-work workflows. The launch comes just days after an internal staff testing build leaked.
The specs haven't changed much: input context tops out at 1 million tokens, the per-turn output cap is 64,000 tokens, knowledge cutoff is March 2026, though some domains are still stuck at January 2025. Distribution rolled out across the board all at once — the Gemini app, AI Studio, the Gemini API, the Gemini Enterprise Agent Platform, Google AI Mode, and the coding tool Antigravity all got access. The model supports adjustable effort levels, letting developers strike their own balance between quality, cost, and latency.
Two Faces in the Same Table
The official comparison table lines up 3.8 Flash alongside 3.7 Flash, Claude Opus 5, Claude Sonnet 5, GPT-5.6 Sol, and GPT-5.6 Terra — the interesting part is the mismatch between price and benchmark scores.
3.8 Flash keeps its promotional pricing of $0.75 per million input tokens and $3.75 per million output tokens; Claude Opus 5 costs $5 and $25, more than six times as much. Despite that price gap, the two are nearly tied on Terminal-bench 2.1 at 89.4 versus 89.1; on CharXiv Reasoning it's 86.2 versus 83.7, on HLE-Verified 54.9 versus 54.4, on the financial-analyst task 61.4 versus 58.6, and on the Harvey legal workflow 10.0 versus 6.7 — in every one of these, 3.8 Flash comes out ahead. On long-video understanding the gap widens further, 87.8 versus 75.4.
Stretch the tasks out, though, and the picture flips. Terminal-bench 4.0, which tests more general agentic ability, gives Opus 5 a score of 51.8 versus just 19.1 for 3.8 Flash; on OSWorld-2.0, which measures an agent operating a computer, it's 75.4 versus 59.0; on the GDPVal-AA v2 Elo score, which gauges knowledge work, it's 1824 versus 1545. On long-horizon software engineering tasks like DeepSWE, it's 74.0 versus 71.0 — a smaller gap, but pointing the same direction.
The pattern is fairly clear: the shorter a task and the closer it is to a single-turn judgment call, the less the price tier matters; the longer it runs, and the more the model has to maintain its own state and correct its own errors, the more the tiers pull apart again. Rough math: running the same medium-complexity coding pipeline, 3.8 Flash's inference cost is roughly a sixth of Opus 5's, with results within a percentage point of each other; swap in an agentic workflow that needs dozens of consecutive steps, and how much of that savings gets eaten up by failed retries is something you'll only know once you've tried it yourself.
The gains over 3.7 Flash, though, are real. DeepSWE climbs from 65.3 to 71.0, Terminal-bench 4.0 from 11.2 to 19.1, OSWorld from 50.6 to 59.0, and the hard tier of biological research workflows jumps from 43.4 to 56.5. For a model that gets a new version every few weeks, that's not a small jump.
Users Knew Before the Announcement
Before the official model card went up, developers in the community had already noticed something off: the model still labeled 3.7 Flash in the web app and in Antigravity would answer "I am 3.8 Flash" when asked what model it was. Others, mid-session, watched a new "3.8 Flash" line suddenly appear in the model picker, with the effort level defaulting to High.
Silently swapping models is standard practice at frontier labs — it lets them test a model under real-world load first and switch back the moment something goes wrong. The cost lands on users: behavior under the same model name changes at some unannounced moment, and prompt effectiveness, token consumption, and output style can all shift along with it, usually without any of it showing up in a changelog. For teams billed by the token, the swing in their bill arrives before the announcement does.
The community's first round of testing relied on the usual tricks: asking the model to generate a solar-system animation in one shot, or draw an SVG of a pelican. Tests like these aren't remotely scientific, but they're low-barrier and the results are obvious at a glance, which usually makes them the only reference point available before official benchmark scores land. Opinions weren't unanimous either — some praised it, some panned it, same as with every model update.
Cheap Has an Expiration Date
The model card lists a few known limitations: occasional slow responses and timeouts persist; when the effort level is turned up, the model burns more tokens to boost performance, so actual cost doesn't necessarily scale linearly with the listed price. On safety, the company says it's roughly on par with 3.7 Flash overall, human red-teaming confirmed it clears the launch bar for child safety, and safety metrics for non-English scenarios saw a slight dip.
Below the pricing column there's a line of fine print: the promotional pricing for 3.7 and 3.8 Flash ends on December 31, and reverts starting January 1 to $1.5 per million input tokens and $7.5 per million output tokens — exactly double. This value-for-money table will need to be recalculated all over again in four months.
Sources: Google DeepMind model card, CocoLoop, Google's official performance comparison table, developer community discussions; context length, output cap, knowledge cutoff, and distribution channels were checked against the model card, while benchmark scores and per-million-token pricing are taken from the official comparison table.