Google released Gemini 3.8 Flash and a specialised cybersecurity variant, Gemini 3.8 Flash Cyber, on 4 September 2026, marking the third Flash-model refresh in six weeks. The company claims the new model delivers measurable gains over its immediate predecessor, Gemini 3.7 Flash, and over rival frontier systems including Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 Sol on multi-step reasoning and software-engineering benchmarks.

Benchmark performance and agentic focus

The headline figure comes from DeepSWE v1.1, a long-horizon software engineering benchmark that tests autonomous, multi-file coding tasks. According to Google, 3.8 Flash outperforms most larger frontier models in autonomously completing these tasks. Sundar Pichai, Google's chief executive, emphasised the model's strength in autonomous problem solving within agentic contexts, workflows where the model plans, executes and iterates over extended sequences without human intervention.

The cybersecurity variant, Flash Cyber, is restricted to selected security firms in Google's Fairwind Program and positions as a counterpart to Anthropic's Claude Mythos 5.1. For the broader developer base, the standard 3.8 Flash is the relevant release, available immediately through Google AI Studio, the Gemini API, Android Studio, the UI development tool Stitch, and for enterprise customers via Gemini Enterprise.

Pricing held flat, with a deadline

Google has kept the per-token price identical to 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens. That undercuts both Claude Opus 5 and GPT-5.6 Sol on list pricing. However, a footnote in the pricing documentation reveals the rate is guaranteed only until 31 December 2026. From 1 January 2027, input tokens rise to $1.50 per million and output tokens to $7.50 per million, a doubling across the board.

The cliff-edge increase means enterprises budgeting for 2027 workloads need to factor in a step change unless Google extends the promotional window or introduces a successor model with revised economics. Given the six-week cadence of Flash releases, another version may well arrive before the price rise takes effect, but no commitment has been made.

Distribution across the Google estate

Rollout is unusually broad for a model launch. 3.8 Flash is already serving queries in Google Search's AI Mode for Google AI Pro and Ultra subscribers, in the Gemini app, and in Google Sheets for the same tiers. Robby Stein, a Google product lead, confirmed the Search integration on X. The speed of deployment reflects Google's advantage in controlling both the model and the distribution surfaces, Search, Workspace, Android tooling and the consumer app, without relying on third-party platforms.

MrBeast partnership targets the creator economy

Alongside the model release, Google announced a high-profile partnership with Jimmy Donaldson, the creator known as MrBeast, whose main YouTube channel exceeds 500 million subscribers. The collaboration spans the Gemini app and Google Health, with Donaldson appearing in a new Gemini marketing campaign and integrating Google products such as Fitbit Air and Gemini tools into his videos. Lorraine Twohill, Google's chief marketing officer, said the first co-produced video would drop on Saturday 6 September.

The deal signals Google's intent to position Gemini as a creative co-pilot for high-volume content producers, not merely a developer tool. By showcasing "wildly impossible ideas" made feasible through AI-assisted ideation and production, Google hopes to drive adoption among the long tail of creators who influence platform trends and, indirectly, advertising inventory.

Competitive context and European relevance

European developers and enterprises evaluating the release face a familiar dynamic: American hyperscalers iterating model generations at a pace that compresses evaluation cycles. The price guarantee until year-end provides a short window for proof-of-concept work at known cost, but the 2027 step-up introduces uncertainty for multi-year contracts. Data residency and GDPR compliance for model inputs processed through Google's US-controlled infrastructure remain separate considerations not addressed in the launch materials.

Anthropic and OpenAI have not yet responded with pricing moves. Claude Opus 5 and GPT-5.6 Sol remain more expensive on published rates, though enterprise discounts and committed-use agreements can narrow the gap. Google's strategy appears to be leveraging its distribution breadth, Search, Android, Workspace, to amortise model training costs across a larger user base than pure API competitors can reach.

People mentioned

  • Sundar Pichai

    Chief executive officer, Google

  • Jimmy Donaldson

    Creator known as MrBeast, YouTube

  • Lorraine Twohill

    Chief marketing officer, Google

Organisations

Google · YouTube