Google Opens Gemini 3.8 Flash to Pro and Ultra
Google has released Gemini 3.8 Flash, calling it the strongest reasoning and programming work model to date, available to Google AI Pro and Ultra subscribers in the Gemini app, Search AI Mode, and spreadsheets, while also entering Gemini Enterprise.
The official positioning shows significant improvements in software engineering, agent tasks, and multi-step reasoning in professional fields compared to the 3.7 Flash released three weeks ago, with speed and costs remaining on par with 3.7. The developer interface introduction price is maintained at $0.75 per million input tokens and $3.75 per million output tokens, with discounts available until December 31. The context window is approximately 1.049 million tokens, with a maximum output of 65,536 tokens, supporting text, image, audio, and video inputs. The knowledge cutoff date is March 2026 in some areas, while others may still be at January 2025.
This is the third Flash version in six weeks, following 3.6 at the end of July and 3.7 three weeks ago. Meanwhile, the cutting-edge Gemini Pro updates have stagnated, with external observers suggesting that the promises of 3.5 Pro may not be fulfilled. The company claims that 3.8 is close to higher-priced cutting-edge models: DeepSWE v1.1 reports 73.7% in long-range software engineering, slightly below Claude Opus 5's 74.0%, and higher than 3.7 Flash's 65.3%; it also lists Vals Finance Agent V2 and Harvey legal agent benchmarks as superior to previous generations. High-effort tiers will consume more tokens.
Simultaneously, Gemini 3.8 Flash Cyber is launched for vulnerability discovery and automatic patching, opened to trusted governments, critical infrastructure, and software maintenance parties through the newly established Fairwind program, replacing 3.5 Flash Cyber. Google claims it exceeds larger parameter commercial models, with the Chrome security team stating its correct patch count is 2.6 times that of larger commercial models; Wiz and Palo Alto Networks provide endorsements. Ordinary users cannot directly access the Cyber version.
Development entry points include Google Antigravity, AI Studio, Android Studio, and Gemini API; free users can try it in AI Studio. The search head stated that 3.8 was released less than three weeks after 3.7. Senior Product Manager Tulsee Doshi and DeepMind Security Head Raluca Ada Popa co-authored a post.
In market mechanisms, buyers are subscribers and enterprises that require low-cost, high-frequency coding and agent workflows, while sellers are cloud vendors using Flash iterations to occupy default entry points. The drive is to push near-cutting-edge capabilities into subscription and API unit prices, rather than releasing another more expensive Pro. Funds flow from Pro/Ultra monthly fees and token bills to Google Cloud inference. Beneficiaries are workflows already tied to Search, spreadsheets, and Studio, while those under pressure are competitors selling similar agents and coding models at higher prices, as well as heavy users still waiting for Pro-level updates.
Source: Public Information
ABAB AI Insight
Google has split Gemini into two product lines, Pro and Flash, betting on Flash for updates by 2026: three releases in six weeks, effectively replacing flagship launches with workhorse models. DeepMind shares the same weight iteration with Search, Cloud, and Chrome security, while carving out a Cyber version for governments and critical infrastructure. This contrasts with OpenAI and Anthropic, which release the largest models first and then smaller versions—Google aims to make 'sufficient and affordable' the default through Search distribution and subscription walls, allowing Pro to be absent. The Antigravity coding environment and embedded spreadsheets integrate the model into existing office entry points rather than creating a separate chat window.
The capital path is to lower prices to lock in volume. The introduction price of $0.75/$3.75 compresses the marginal cost of agent cycles to a daily usable level, while recovering fixed costs through Pro and Ultra memberships. Fairwind turns the strongest vulnerability discovery capabilities into controlled exports, preventing equivalent tools from reaching unvetted users, while securing defense and critical infrastructure contracts. The motivation is that token consumption for coding and agent tasks far exceeds that for Q&A; whoever can offer cheaper at equivalent scores will capture the enterprise default gateway. Strategically, Flash approaches cutting-edge scores, allowing Pro to delay and convert training budgets into a denser release calendar.
Analogous examples include AWS's use of Graviton and Spot to capture computing, Microsoft embedding Copilot into Office, and Anthropic using mid-sized models for programming agents. Google is in the phase of replacing flagship narratives with work models: distribution is present, but the unit cost to migrate enterprise pipelines from Claude/GPT is lacking. The Cyber version separates security capabilities from general chat, turning them into licensed products.
This is a technological substitution. What is being replaced is not the 'larger model' but the pricing structure that mandates purchasing the most expensive model to perform agent tasks. The mechanism is that once the benchmark gap narrows to a few percentage points, latency, price, and whether it is integrated into Search or spreadsheets will determine migration; high-frequency scenarios will abandon flagship premiums. Flash releases every three weeks make substitution a schedule rather than a one-time launch.
ABAB News · Cognitive Laws
Fast iteration of work models allows flagship to be delayed.
Once scores are close, entry points determine migration.
The strongest capabilities are often sold first to controlled customers.