Google launches Gemini 3.8 Flash at unchanged prices, but tasks cost more tokens

What shipped

Google launched Gemini 3.8 Flash on 2 September 2026, three weeks after Gemini 3.7 Flash and six weeks after 3.6 Flash, according to two comparison write-ups published on dev.to. The new model arrives with per-token pricing identical to its predecessor: $0.75 per million input tokens and $3.75 per million output tokens, thinking included, holding until 31 December 2026, when the standard rate of $1.50 and $7.50 takes effect. The model ID is gemini-3.8-flash, and Google describes the release as matching 3.7 Flash on both price and speed.

What carried over

Most of the specification is unchanged. Input is capped at 1,048,576 tokens with 65,536 output tokens; accepted inputs remain text, image, video, audio and PDF, with text-only output. Context cache reads cost $0.075 per million tokens at launch, batch processing keeps its 50% discount, and the 5,000 free monthly Google Search grounding requests are shared across the Gemini 3.x family. Both the Interactions API and the legacy generateContent endpoint remain fully supported.

Speed is close to the older model. Artificial Analysis measured 302.1 output tokens per second at the high thinking level, but time to first token was 13.3 seconds because the model reasons before it starts writing an answer.

What actually changed

The difference is behavioural. Google positions 3.8 Flash as its most capable Flash model, aimed at long-horizon software engineering, autonomous agents and complex enterprise workflows. On difficult tasks it takes more reasoning steps, splits them into smaller units, calls tools repeatedly and verifies its own output mid-run.

Google's published benchmark deltas over 3.7 Flash are modest but consistent: Vals Finance Agent v2 at 61.4% versus 59.0%, HLE-Verified at 54.9% versus 53.6%, and the Harvey legal agent benchmark at 10.0% versus 8.8%. In all three, 3.8 Flash also outscored more expensive frontier models in Google's comparison table, including Claude Opus 5 and GPT-5.6 Sol.

Artificial Analysis gives 3.8 Flash an Intelligence Index of 59 at high effort, up from 56 for 3.7 Flash and 52 for 3.6 Flash. On its τ³-Banking tool-use benchmark the new model reaches 45%, a 12-point jump that may matter more for agentic workloads than the headline index. Google separately claims 3.8 Flash completes more than three times as many tasks as 3.7 Flash in long, document-heavy workflows, but that is an internal evaluation with no public harness, so treat it as directional.

Coding data is incomplete. DeepSWE v1.1 scores of 65.3% for 3.7 Flash and 49.0% for 3.6 Flash are on record, while 3.8 Flash's exact figure appears only in a chart, and no text figures are published for SWE-Bench Pro, Terminal-Bench or OSWorld.

On safety, Google reports a significant jump on Gray Swan prompt injection tests without publishing an overall score. Compared with 3.7 Flash, multilingual safety rose 5.4 points, text-to-text safety dipped 0.4, and unjustified refusals rose 1.1, hinting at slightly more over-refusal.

Same token price, higher task bill

Per-token costs are flat, but consumption is not. Artificial Analysis measured roughly 48,000 output tokens per task on average for 3.8 Flash, about 30% more than 3.7 Flash. On a per-task basis, 3.7 Flash at high cost $0.40 over 2.2 minutes, while 3.8 Flash costs $0.24 at low effort (0.8 minutes), $0.41 at medium and $0.58 at high (2.5 minutes). High-effort 3.8 is therefore about 45% more expensive per task with 14% more wall-clock time, medium roughly matches 3.7 at high, and low is the cheapest and fastest option for latency-sensitive paths. The thinking level is the main dial for trading quality against cost and latency.

Two breaking changes

First, the minimal thinking level is gone. 3.8 Flash accepts only low, medium (the default) and high; configurations using minimal now fail validation and should be mapped to low before migrating.

Second, function results must include both call_id and name in the Interactions API, and functionResponse on generateContent must likewise return the matching identifier and function name. Systems that key results only by function name can break on the second turn of a tool loop.

Google's migration guide also restates existing Gemini 3 rules worth rechecking: keep temperature at the default 1.0, use thinking_level rather than an integer thinking_budget, drop candidate_count, and return thought signatures exactly as received.

Why it matters

Google says 3.7 Flash remains fully supported with no deprecation date announced, and 3.8 Flash has a March 2026 knowledge cutoff, so migration is not urgent. The real story is the pricing pattern: flat token rates paired with heavier per-task consumption shift the cost conversation from rates to workload. Teams running multi-step agents should benchmark against their own traffic before switching, and the two API changes mean even a simple model-string swap can fail in production if it goes untested.