Google Launches Gemini 3.8 Flash at a Fraction of Claude Opus 5's Price, Benchmarks Split Down the Middle

Google DeepMind released Gemini 3.8 Flash and a cybersecurity-focused sibling, Gemini 3.8 Flash Cyber, on Wednesday, September 2, according to a post from Google's own blog credited to product management senior director Tulsee Doshi and DeepMind security lead Raluca Ada Popa. It's the third Flash-tier model Google has shipped in six weeks, following 3.6 Flash and 3.5 Flash-Lite in July and 3.7 Flash on August 13, per 9to5Google.

The pitch is simple: cheaper, and in a lot of cases, just as good. Gemini 3.8 Flash runs $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026, after which the price jumps to $1.50 and $7.50 starting January 1, 2027, according to both 9to5Google and The Decoder. Claude Opus 5, which Anthropic launched July 24, costs $5 per million input tokens and $25 per million output, per multiple outlets including Office Chai. GPT-5.6 Sol from OpenAI sits in between at $4 input / $20 output.

Where Gemini wins

On a batch of practical, agent-heavy benchmarks, Gemini 3.8 Flash actually beats the pricier models. Office Chai reported it scores 61.4% on Vals Finance Agent v2 versus Opus 5's 58.6% and GPT-5.6 Sol's 53.8%. On Harvey's Legal Agent Benchmark it hits a 10.0% pass rate, more than double GPT-5.6 Sol's 2.5% and well ahead of Opus 5's 6.7%. It edges Opus 5 on Terminal-bench 2.1 (89.4% to 89.1%), leads on CharXiv Reasoning (86.2% to 83.7%), and posts 54.9% on HLE-Verified, a multidisciplinary reasoning test, narrowly ahead of Opus 5's 54.4%.

Google's own blog post confirms the HLE-Verified score and says the model's gains come from a design choice. 3.8 Flash "works harder," running extra reasoning steps and calling tools iteratively on complex tasks, at the cost of burning more tokens.

Where Opus 5 still wins, and by how much

Opus 5 leads by a wide margin on GDPVal-AA v2, an Elo-scored knowledge-work benchmark, at 1824 versus Gemini's 1545, according to Office Chai. The gap widens further on Terminal-bench 4.0, a general agent capabilities test, where Opus 5 scored 51.8% against Gemini's 19.1%. Opus 5 also leads on OSWorld-2.0, a computer-use benchmark, 75.4% to 59.0%.

On the long-horizon software engineering benchmark DeepSWE v1.1, the numbers actually diverge across outlets. Office Chai listed Opus 5 at 74.0% and Gemini 3.8 Flash at 71.0%. The Decoder, citing Google's own figures, put Gemini 3.8 Flash at 73.7% against Opus 5's 74.0%, a near-dead heat. 9to5Google's writeup flagged an "updated DeepSWE v1.1 score," suggesting the number moved after initial publication. Crypto Briefing's framing, that Gemini 3.8 Flash "beats larger frontier models on engineering tasks," glosses over the fact that on this specific benchmark, under either version of the numbers, Opus 5 still comes out on top.

DeepMind's new head, Koray Kavukcuoglu, said Google isn't only chasing price-performance and still wants to lead on raw capability, according to The Decoder. That comment lands against a backdrop The Decoder also flagged: Google's flagship frontier models, Gemini 3.5 Pro and Gemini 4, remain unreleased while the company ships a third Flash update in six weeks.

Artificial Analysis, an independent benchmarking outfit, gave Gemini 3.8 Flash an Intelligence Index score of 59, three points above 3.7 Flash's 56, putting it roughly level with GPT-5.6 Sol at high reasoning settings and Grok 4.6 at medium settings, per The Decoder. On cost efficiency, Artificial Analysis put the model on the "Pareto frontier" at $0.58 per task.

Google is already defaulting to Gemini 3.8 Flash in its Antigravity coding tool and Managed Agents product, according to daily.dev, and the model was reportedly visible in Google Cloud's Agent Studio before the official rollout. It's live now for Google AI Pro and Ultra subscribers, in AI Mode, Google Sheets, and through the Gemini API.

Explainx.ai had argued earlier in the week that no Google-confirmed release existed yet and that reports of a Wednesday launch traced back to a Wall Street Journal report and leak sites describing an internal codename, "skimaki." Google's own blog post, dated September 2, settles that question: the model shipped as reported, with a published model card and benchmark table, not just internal engineer chatter.

The open question is what happens to that DeepSWE discrepancy and to the rest of Google's benchmark claims once third-party evaluators outside Artificial Analysis run their own numbers. Google's pricing math also has a clock on it. The $0.75/$3.75 rate doubles on January 1, 2027, which gives enterprise customers roughly four months to lock in the cheap tier before the real cost of running Flash at scale kicks in.