Originally published at IndieSeek.
Gemini 3.8 Flash is Antigravity's default: audit agent cost and tool loops before rollout
Quick answer
Google released gemini-3.8-flash as a generally available Gemini API model on September 2, 2026. It supports a 1,048,576-token input window, up to 65,536 output tokens, multimodal input, function calling, structured output, code execution, caching, and Computer Use in preview. Google also changed the Antigravity managed agent and its SDK to use Gemini 3.8 Flash by default.
That default change is the operational risk. An unchanged agent identifier can now produce a different number of reasoning steps, tool calls, tokens, and retries. Google explicitly says 3.8 Flash may use more tokens on long, complex tasks by design. Before accepting the new default, record the effective model configuration, set a thinking level per route, cap the loop, and compare accepted-run cost against an explicitly pinned 3.7 baseline.
Who this is for
This guide is for indie developers and platform teams using Gemini Managed Agents, the Antigravity SDK, or a direct Gemini API agent loop. It focuses on a default-model change, not a generic benchmark contest.
If you are migrating direct requests and need a broad model comparison, start with the Gemini 3.7 vs 3.6 migration checklist. For high-volume routine work, the Flash vs Flash-Lite routing guide remains the cheaper-task decision layer.
What changed—and what did not
| Boundary | Confirmed state | Operator consequence |
|---|---|---|
| API model |
gemini-3.8-flash is GA |
Pin this exact ID for a direct-API canary |
| Managed agent | Antigravity now uses 3.8 Flash by default | Re-test even when the agent name did not change |
| Thinking |
low, medium, high; medium is default |
Configure it explicitly; minimal returns an error |
| Long tasks | More reasoning and iterative verification can consume more tokens | Budget the whole run, not one request |
| Standard price | $0.75 input / $3.75 output per 1M tokens through 2026-12-31 | Same list rate as 3.7 does not guarantee the same run cost |
| 2027 price | $1.50 input / $7.50 output from 2027-01-01 | Recalculate every production budget before the date boundary |
| Tool surface | Function calling, code execution, file/search tools, and preview Computer Use are supported | Re-run tool contracts and safety gates |
| Live API | Not supported | Keep real-time voice or streaming routes on a supported model |
Output pricing includes thinking tokens. A run with 100,000 uncached input tokens and 20,000 output tokens, including thinking, costs about $0.15 at the introductory Standard rate and $0.30 from January 1, 2027, before grounding or cache-storage charges. The formula is simple; predicting the token counts is not.
Separate identity, effort, and authority
Treat the rollout as three independent controls:
- Identity: which model the direct request or managed-agent configuration resolves to.
- Effort: which thinking level, token usage, retries, and tool rounds the run consumes.
- Authority: which tools, files, network destinations, and irreversible actions the agent can reach.
A stronger model does not reduce the need for a sandbox or approval boundary. Keep untrusted repositories behind the AI coding-agent sandbox checklist, especially when Computer Use is enabled.
Build a shadow-run ledger
Choose five representative tasks: one focused code fix, one multi-file refactor, one search-and-summarize job, one recoverable tool failure, and one task that must stop before an external write. Run each against the Antigravity default and a 3.7-pinned control where the supported configuration allows it.
Keep the task fixture, repository revision, instructions, tools, network policy, timeout, and acceptance tests identical. Record this per run:
agent_id: antigravity
configured_model: default | gemini-3.7-flash | gemini-3.8-flash
observed_at: 2026-09-03T18:00:00+08:00
thinking_level: low | medium | high
input_tokens:
output_and_thinking_tokens:
cached_input_tokens:
tool_calls:
repeated_tool_calls:
wall_time_seconds:
accepted: true | false
human_repair_minutes:
side_effect_attempted: false
If the runtime does not expose a resolved model field, do not invent one. Preserve the configured agent_config, request timestamp, SDK version, and response metadata that is actually available, and label the effective model as inferred from Google's documented default.
A six-gate rollout
1. Freeze the baseline
Save the current agent configuration, Gemini SDK version, tool schemas, safety policy, retry policy, and five-task fixture revision. Keep the 3.7 route available for rollback.
2. Make thinking explicit
Use low for latency-sensitive classification or retrieval, medium for ordinary coding and agent tasks, and high only for the hardest multi-step work. Add a negative test proving minimal fails rather than silently mapping it.
3. Cap the loop
Set maximum wall time, model turns, tool calls, repeated identical calls, and total token or spend allowance. Stop when the agent repeats the same failed action, not when the account-wide budget is exhausted.
4. Verify tool contracts
Test one valid tool call, one schema error, one timeout, one permission denial, and one irreversible action that requires a human. Computer Use remains preview; do not let a model upgrade convert a read-only canary into a production desktop session.
5. Compare accepted-run economics
Use accepted_run_cost = model cost + retry cost + human repair cost. Because 3.8 and 3.7 share the same current list rates, a higher 3.8 token count may still win if it removes retries—or lose if extra verification adds no accepted result.
6. Promote with a dated budget
Ramp one low-risk route, then 5%, 25%, and 100%. Require acceptance, latency, repeated-call rate, policy compliance, and cost to stay inside limits. Store both the current rate and the January 1 rate in the decision record so the rollout cannot rely on promotional pricing forever.
Canary matrix
| Canary | Pass condition |
|---|---|
| Default identity | Configuration and timestamp prove whether default or pinned routing was used |
minimal thinking |
Request fails visibly; router does not retry forever |
| Recoverable tool error | Agent corrects once and completes inside the call cap |
| Repeated tool loop | Circuit breaker stops the run with a clear reason |
| Irreversible action | Human approval is required before the side effect |
| Live route | Unsupported Live API path stays on its existing supported model |
| Cost boundary | Ledger calculates both 2026 and 2027 rates |
| Rollback | Pinned 3.7 control can still complete the baseline fixture |
Copyable decision record
owner / date / route:
agent_id / sdk_version / configured_model:
baseline_model / candidate_model:
fixture_revision / tool_schema_revision:
thinking_levels_tested:
accepted_runs / total_runs:
input / cached / output-plus-thinking tokens:
tool_calls / repeated_calls / retries:
intro_rate_cost / 2027_rate_cost:
human_repair_minutes / policy_violations:
promotion: hold | 5% | 25% | 100% | rollback
rollback_model / rollback_trigger:
Building something? Take a 60-sec game break. Score to rank your product or profile on tapto.top and get more exposure—free, no signup.
Common mistakes
- Assuming an unchanged Antigravity agent ID means the underlying model did not change.
- Comparing 3.8
highwith 3.7mediumand calling the result a model improvement. - Looking only at per-token price while ignoring thinking tokens, retries, and tool loops.
- Sending
minimal, then hiding the validation error behind unbounded retries. - Treating Computer Use preview as permission to remove human approval.
- Forgetting that the published Standard rate doubles on January 1, 2027.
FAQ
Is Gemini 3.8 Flash generally available?
Yes. Google's September 2 Gemini API release notes identify gemini-3.8-flash as GA.
Is 3.8 more expensive than 3.7 per token?
Their published Standard, Batch, Flex, Priority, caching, and storage rates currently match. Actual run cost can differ because token use, tool calls, retries, and acceptance can differ.
Can I keep Antigravity on 3.7?
Google's Agents documentation says the underlying model can be configured, including Gemini 3.7 Flash. Verify the currently supported agent_config shape in your SDK and keep the pin as a rollback control.
Does Gemini 3.8 Flash support the Live API?
No. The model page lists Live API as unsupported. Route real-time workloads separately.
Top comments (0)