AI NewsSep 3, 2026Words 1608Read time≈ 5 min

Google Launches Gemini 3.8 Flash for Long-Horizon Coding and Agents

Gemini 3.8 Flash adds stronger coding and agentic reasoning, a 1M-token context window, adjustable effort, and introductory API pricing.

AIGeminiAgentTool Use

Contents · 13

1. Google ships its third Flash update in six weeks
2. Longer reasoning loops are the central product change
3. Google reports its largest gains in coding and agent evaluations
4. Migration requires more than changing the model name
5. Pricing is unchanged initially but doubles in January
6. The model card preserves important limitations
Frequently Asked Questions
When was Gemini 3.8 Flash released?
What is the Gemini 3.8 Flash API model ID?
Does Gemini 3.8 Flash support images, audio, and video?
How much does Gemini 3.8 Flash cost?
Is Gemini 3.8 Flash Cyber generally available?
Sources

1. Google ships its third Flash update in six weeks

Google released Gemini 3.8 Flash on September 2, positioning the model as its most capable Flash-tier system for software engineering, autonomous agents, and multi-step professional work. The stable API identifier is gemini-3.8-flash, and the model launched at general availability rather than as a preview.

The release follows Gemini 3.7 Flash, introduced on August 13, and Gemini 3.6 Flash, released on July 21. That makes Gemini 3.8 Flash Google’s third Flash release in six weeks. Google says the new model retains the speed and introductory price of 3.7 Flash while improving its ability to sustain reasoning and tool use over longer tasks.

Gemini 3.8 Flash accepts text, images, audio, video, and PDF documents. Its context window is 1,048,576 input tokens, with a maximum output of 65,536 tokens. Output remains text-only: the model can analyze multimodal material but does not generate images or audio.

The model supports function calling, file search, code execution, structured output, URL context, Google Search and Maps grounding, and context caching. Computer use is supported in preview. The Live API, image generation, and audio generation are not supported.

Google announced a second model, Gemini 3.8 Flash Cyber, alongside the general-purpose release. The two models share the same underlying intelligence, according to the company, but the Cyber version has more permissive cybersecurity capabilities and is limited to vetted governments, critical-infrastructure operators, software maintainers, and other trusted defenders through Google DeepMind’s Fairwind Program.

2. Longer reasoning loops are the central product change

The defining change from Gemini 3.7 Flash is not a larger context window or a new output modality. Those core limits remain unchanged. Instead, Google says 3.8 Flash is more willing to perform additional reasoning steps, recover from obstacles, and call tools repeatedly when a task cannot be completed reliably in one pass.

This behavior is intended for work such as navigating a large codebase, diagnosing failures, editing multiple files, testing the result, and revising the implementation. Google demonstrated the model in Antigravity building a 3D game, playing it to identify errors, and applying code changes within a continuing agent loop. Other demonstrations included a DOS-style Google Maps application, a data-backed topographic visualizer, and a Three.js hardware-teardown tool.

Those demonstrations are controlled examples rather than independent product tests. Their relevance lies in the workflow: Gemini 3.8 Flash is designed to alternate between reasoning, tool calls, execution, observation, and revision instead of merely returning a code sample.

The added diligence has a measurable cost. Google explicitly warns that 3.8 Flash can consume more tokens than 3.7 Flash, particularly at higher reasoning settings. Developers can select LOW, MEDIUM, or HIGH through the thinking_level parameter; MEDIUM is the default. The model does not accept a MINIMAL level, and requesting it produces a validation error.

For latency-sensitive or high-volume applications, Google recommends lowering the effort level or continuing to use Gemini 3.7 Flash. The practical choice is therefore workload-dependent: 3.8 Flash offers more persistence for difficult tasks, but 3.7 Flash remains supported for deployments where predictable token consumption matters more than maximum task completion.

3. Google reports its largest gains in coding and agent evaluations

Google’s published evaluations show improvements across several coding and tool-oriented tests, although the results remain vendor-reported and benchmark performance does not establish reliability in a specific production environment.

On Terminal-bench 2.1, Gemini 3.8 Flash scored 90.8%, compared with 81.6% for Gemini 3.7 Flash. Its SWE-Atlas score increased from 48.0% to 51.9%, while SWE-Bench Pro moved from 60.4% to 61.6%. On the banking subset of τ³-bench, which tests agents handling multi-step interactions, the reported score rose from 30.9% to 38.1%.

The model also reached 61.4% on Vals Finance Agent v2, ahead of Gemini 3.7 Flash at 59.0% in Google’s comparison. On Harvey’s Legal Agent Benchmark, the respective scores were 10.0% and 8.8%.

For HLE-Verified, an expert-level multidisciplinary reasoning benchmark, Google reported 54.9% for Gemini 3.8 Flash and 53.6% for 3.7 Flash. The published comparison placed 3.8 Flash narrowly above GPT-5.6 Sol at 54.5% and Claude Opus 5 at 54.4%. Small differences of this size should not be treated as proof that one model will consistently outperform another across unrelated workloads.

Google also says Gemini 3.8 Flash leads the DeepSWE v1.1 long-horizon software-engineering evaluation while operating at lower cost than larger models. Ars Technica’s review of the release characterized the gains as modest on many general tests but larger on coding evaluations. It also noted that, despite improving over 3.7 Flash on OSWorld 2.0, the new model remained well behind Claude Opus on that computer-use benchmark.

The mixed pattern matters for deployment decisions. The strongest published case for upgrading concerns terminal work, coding agents, and extended task execution—not a uniform jump across every reasoning or computer-control test.

4. Migration requires more than changing the model name

Developers moving from Gemini 3.7 Flash or 3.6 Flash must change the model identifier to gemini-3.8-flash. Google’s enterprise migration guide also directs users to replace the older integer-based thinking_budget configuration with the thinking_level enum.

Several legacy generation controls are unavailable. Google says temperature, top_p, and top_k are ignored by the backend, while frequency_penalty, presence_penalty, and candidate_count produce API errors. Applications that depend on these fields need configuration changes rather than a model-name substitution alone.

Function calling is more strictly validated. A function-response turn must match the preceding function call’s identifier, name, and execution count. Multimodal assets must be placed inside the response payload, chat histories cannot end with a model-role turn, and prefilled model responses are unsupported.

These constraints can expose integration errors that older deployments tolerated. Teams evaluating the upgrade should test complete agent traces—including retries, tool responses, structured outputs, and conversation-history reconstruction—rather than comparing isolated prompts.

The model’s documented knowledge cutoff is March 2026, although Google cautions that coverage in some domains may still resemble the January 2025 cutoff associated with the broader Gemini 3 family. Search grounding or another retrieval layer remains necessary for current or rapidly changing information.

5. Pricing is unchanged initially but doubles in January

Through December 31, 2026, standard Gemini API usage costs $0.75 per million input tokens and $3.75 per million output tokens, including thinking tokens. These are the same introductory rates listed for Gemini 3.7 Flash.

The rates double on January 1, 2027, to $1.50 per million input tokens and $7.50 per million output tokens. Context-caching charges likewise rise from $0.075 to $0.15 per million tokens, while cache storage increases from $0.50 to $1 per million tokens per hour.

Batch and Flex inference cost $0.375 per million input tokens and $1.875 per million output tokens during the introductory period. Both increase to $0.75 and $3.75 respectively in January. Priority inference is priced higher, at $1.35 per million input tokens and $6.75 per million output tokens through December.

The output price includes internal thinking tokens, making effort selection part of the cost model. A task that invokes more reasoning steps can cost more even when its final visible response is short. Production evaluations should therefore record total billed tokens, completion rate, latency, and tool-call count rather than comparing list prices alone.

Developers can access Gemini 3.8 Flash through Google AI Studio, the Gemini API, Android Studio, Stitch, and Antigravity. Enterprises can use it through the Gemini Enterprise Agent Platform. Consumers need a Google AI Pro or Ultra subscription to access it in the Gemini app, AI Mode in Google Search, and Gemini in Google Sheets.

6. The model card preserves important limitations

Google’s model card says Gemini 3.8 Flash can hallucinate, occasionally respond slowly, time out, or consume more tokens than expected at higher effort levels. Its multimodal input support and agent tools do not remove the need to validate generated code, financial analysis, legal work, or actions taken through external systems.

The company reports similar or improved general content-safety performance relative to Gemini 3.7 Flash, although automated testing found a slight regression in some non-English safety evaluations. Google says manual review determined that the flagged losses were predominantly false positives or non-egregious cases.

Under Google DeepMind’s Frontier Safety Framework, the company concluded that Gemini 3.8 Flash does not add material capabilities over 3.7 Flash in the framework’s tracked high-risk domains and is unlikely to cross a Tracked or Critical Capability Level. The unrestricted model retains safeguards covering cyber offense and chemical, biological, radiological, and nuclear misuse; the more capable Cyber variant is distributed separately under restricted access.

Frequently Asked Questions

When was Gemini 3.8 Flash released?

Google released the stable model on September 2, 2026, three weeks after Gemini 3.7 Flash.

What is the Gemini 3.8 Flash API model ID?

The stable API identifier is gemini-3.8-flash.

Does Gemini 3.8 Flash support images, audio, and video?

It accepts text, images, audio, video, and PDFs as input but produces text only.

How much does Gemini 3.8 Flash cost?

Standard API pricing is $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. The rates rise to $1.50 and $7.50 on January 1, 2027.

Is Gemini 3.8 Flash Cyber generally available?

No. Google limits the Cyber variant to approved defenders through the Fairwind Program. The general-purpose Gemini 3.8 Flash model is generally available.