Google is back in the top tier of the AI race. Gemini 3.8 Flash, released on September 2, 2026, scores 74% on the DeepSWE v1.1 benchmark for software engineering, tying GPT-6 Astra and Claude Opus 5. It does that at an average cost of $2.36 per task against $11.84 for Opus 5.
On raw intelligence, Claude Fable 5.1 is still ahead. On useful work per dollar, Gemini 3.8 Flash beats everything near its capability level, and for most production workloads that second number is the one that decides your bill.
At TJ Digital we run around 50 prompts per client through the big models to see who gets recommended, and every workflow behind that work uses at least two models, usually three. This release did not change that setup. It reinforced why we built it that way.
The release calendar is the reason. Anthropic shipped Claude Fable 5.1 on September 1 and Google shipped 3.8 Flash on September 2. Z.ai released the open weight GLM-5.3 on August 14, one day after Google’s own 3.7 Flash.
So a company that picked “the best model” in mid August had three credible challengers show up inside of a month. That happens again next quarter, and the quarter after that.
What Is Gemini 3.8 Flash Actually Good At?
Gemini 3.8 Flash is Google’s workhorse model for coding and agents. It has a 1 million token context window, takes text, image, video and audio input, and exposes low, medium and high reasoning levels. Google also made it the default model behind its Antigravity managed agent.
“Flash” no longer means a small model for small jobs. Google built 3.8 to take more reasoning steps, call tools repeatedly and check its own work, which is why it can burn through far more tokens than 3.7 on a hard task.
How Does It Compare on Coding Benchmarks?
The DeepSWE v1.1 leaderboard is the most useful number here, because every model was run through the same agent scaffold instead of each vendor’s own setup. Here is where 3.8 Flash landed on the September 3 board.
| Model | DeepSWE v1.1 | Avg. task cost | Output tokens | Agent steps |
| GPT-6 Astra, xhigh | 74% ±3 | $6.52 | 30k | 29 |
| Gemini 3.8 Flash, high | 74% ±1 | $2.36 | 143k | 166 |
| Claude Opus 5, max | 74% ±4 | $11.84 | 118k | 99 |
| GPT-5.6 Sol, max | 73% ±3 | $6.46 | 60k | 61 |
| Claude Fable 5, xhigh | 70% ±3 | $13.41 | 80k | 68 |
| GLM-5.3, max | 69% ±3 | $3.99 | 80k | 124 |
| Gemini 3.7 Flash | 65% ±3 | $2.03 | 94k | 117 |
Look at the token column before you get too excited. Flash needed 143,000 output tokens and 166 agent steps to get there, against 29 steps for Astra. It reaches the same score partly by grinding, which Google can afford because inference on this model is cheap and fast.
That matters if your tool calls are slow, rate limited or billed separately. In that case, some of the price advantage disappears into the work around the model.
Is Flash Actually the Fastest Model?
It is fast at generating tokens. Artificial Analysis measured roughly 300 output tokens per second at high reasoning around launch, with current measurement near 278.
Tokens per second is not the same thing as finishing a task quickly. High reasoning 3.8 produced about 30% more output tokens per task than 3.7, which pushed measured cost per task from about $0.40 to $0.58 and task time from 2.2 minutes to 2.5. Fable 5.1 at medium reasoning finished the equivalent workload in about 2.1 minutes.
So a fast model does not automatically give you a fast agent. Agent speed depends on how many turns, tool calls and verification loops the model decides to run.
How Does Flash Handle Legal and Financial Work?
This surprised me more than the coding results. Google reports 61.4% on the Vals Finance Agent v2 benchmark against 58.6% for Claude Opus 5, and 10.0% all-pass on Harvey’s legal agent benchmark against 6.7% for Opus 5.
The Harvey benchmark is not legal trivia. It asks an agent to answer a client inquiry using shell access, file editing, and Word, Excel and PowerPoint tools, so an all-pass means the model produced an entire piece of legal work product.
Those numbers come from Google’s own comparison table, so treat them as directional rather than a clean head to head. Google also reports 54.9% on HLE-Verified, which puts it near models that cost many times more.
How Much Does Gemini 3.8 Flash Cost Compared to Claude?
Through December 31, 2026, Gemini 3.8 Flash runs $0.75 per million input tokens and $3.75 per million output tokens. On January 1, 2027 that doubles to $1.50 and $7.50.
Claude Fable 5.1 lists at $10 per million input and $50 per million output. At the introductory rate, Google is about 13 times cheaper on both sides, and still about 6.7 times cheaper once the discount ends. Anthropic narrows the real gap with prompt caching, where cache reads cost $0.25 per million tokens and cut typical workload costs by around 25%, or up to roughly 45% on heavily agentic work.
Track cost per completed task. Price per token will mislead you, and Gemini proved it with this release. Per-token pricing did not change from 3.7 to 3.8, and cost per high reasoning task still went up about 40% because the model does more work.
Where Is Claude Fable 5.1 Still Ahead?
Fable 5.1 is built for jobs that run for hours across multiple applications, and the hardest benchmarks still show a real gap.
Anthropic reports 55.8% for Fable 5.1 on Terminal-Bench 4. Google’s comparison table lists 19.1% for Gemini 3.8 Flash on the same benchmark. Different labs ran those tests, so the exact spread is arguable, but the size of it is hard to wave away.
The pattern repeats on broader professional work. Fable 5.1 scores 1,853 on GDPval-AA v2 against 1,545 for 3.8 Flash. Flash is strong at specific jobs and weaker across knowledge work more broadly, which is what I expected after using it.
Here is the simplest way I can frame the three-way choice right now.
| Gemini 3.8 Flash | Claude Fable 5.1 | GLM-5.3 | |
| Best at | High volume coding and agents | Long, hard, multi-app work | Control and portability |
| Input / output price | $0.75 / $3.75 (intro) | $10 / $50 | Varies by deployment |
| DeepSWE v1.1 | 74% ±1 | Fable 5 scored 70% ±3 | 69% ±3 |
| Modalities | Text, image, video, speech | Text and image | Text only |
| Main limitation | Token hungry, trails on the hardest agentic tasks | Expensive | Below the top closed models |
GLM-5.3 deserves more attention than it gets. It sits within five points of Flash on the same leaderboard, ships open weights, and speaks both OpenAI and Anthropic compatible protocols. That last part is a quiet admission from Z.ai that developers want the ability to move between ecosystems without rewriting their application.
Why Is Google Investing So Heavily in Flash Models?
Almost all of Google’s revenue comes from Search ads, and Search is becoming AI Search. That explains the Flash strategy better than any benchmark does.
At I/O in May, Google said AI Mode had passed one billion monthly users and that AI Mode queries were more than doubling every quarter. It made Gemini 3.5 Flash the global default model for AI Mode at the same time. Alphabet’s second quarter disclosure put Gemini models at roughly 22 billion API tokens per minute.
An AI answer costs far more to produce than a ranked list of links. Once AI is the default Search experience, inference efficiency stops being an API optimization and becomes a cost line at Google scale.
You can see the strategy in the release order.
- 3.6 Flash cut output tokens by 17% against 3.5.
- 3.7 Flash raised capability and halved the price.
- 3.8 Flash kept the price and spent the savings on more reasoning for hard tasks.
Google is optimizing useful work per dollar and per second. It can also spread one Flash model across Search, Workspace, Cloud and its developer agents, which no competitor can match on distribution.
Sundar Pichai has told Reuters that large companies could save more than $1 billion a year by moving most AI workloads to Gemini. Treat that as a sales number, but it shows how central price performance is to the pitch.
What Happened to Gemini 3.5 Pro?
It is late. In July, Google said 3.5 Pro was still testing with partners and would ship when it was ready, and in the same announcement confirmed it had already started its most ambitious pre-training run yet for Gemini 4.
My read is that they will skip straight to Gemini 4 and make that the real answer to Claude and OpenAI. Google has not said that, so treat it as my guess rather than a fact.
What is confirmed is the reorganization. On August 5, Demis Hassabis moved into the chairman role at Google DeepMind plus a new Alphabet chief scientist position, with Koray Kavukcuoglu taking day-to-day responsibility for DeepMind. Around the same time, Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le left to start Discovery Loop.
Google has clearly not given up on competing at the top. It feels like they have solved something internally on the Flash line that they have not yet solved on the frontier line.
Why Should You Avoid Locking Into One AI Model?
Seven weeks, three Flash releases, plus Fable 5.1 and GLM-5.3. That release calendar is the whole argument.
Lock-in shows up in more places than people expect.
- Capability lock-in. The best model for your workload can change before your migration finishes.
- Economic lock-in. Prices and token consumption both move. Gemini held its price and still got 40% more expensive per task.
- API lock-in. Model generations change parameters. GLM-5.3 removed the ability to disable reasoning, so applications using that setting fail until they are reconfigured.
- Workflow lock-in. If your tool definitions, retrieval logic and business rules live inside one provider’s agent API, switching means rebuilding.
- Knowledge lock-in. This is the dangerous one. If your company’s context only exists inside a provider’s proprietary memory or vector store, you can replace the model but not the application.
- Observability lock-in. Without comparable cost, latency and success data across providers, you cannot prove a challenger is better.
None of this means you avoid proprietary features. Sometimes a specific cache, search implementation or managed agent is the reason a model wins. Make that dependency a decision you made on purpose, with a number attached, instead of something that happened to you.
How Do You Build Model Agnostic AI Workflows?
The durable assets are your knowledge, your tools, your evaluations and your workflows. The model underneath should be replaceable. Here is the order I would build in.
Start with the knowledge base. Keep your source documents, permissions, metadata and lineage in your own systems, outside any model platform. If embeddings change, you want to re-embed the corpus rather than rebuild it from opaque provider IDs. This is the same reason we build a Brand Ambassador for every client, a documented knowledge base about the business that any model can read.
Put a gateway between your apps and your providers. Your application should ask for something like “analyze this contract, high quality, 30 second budget” instead of hard-coding a model name in fifty places. LiteLLM is one open source version of this pattern. Model selection then becomes a config change instead of a release.
Define your tools once. MCP exists to standardize the boundary between an AI application and your data and tools, so the same server can serve different model clients. Writing separate integrations per provider is how you end up stuck.
Build an evaluation set from your real work. Public benchmarks screen candidates. They do not tell you whether a model follows your return policy or your intake procedure. Track task success, reliability across repeated runs, tool use, cost per completed task, end to end latency and how much human correction each run needs. NIST’s AI risk management guidance makes the same point about testing and validation rather than trusting vendor claims.
Run champion and challenger. Keep your production model as the champion, run new releases against your own evaluation set, and promote only when the improvement clears a threshold you set in advance. Watch models weekly. Migrate when your own evidence says to.
Route by task. A Flash class model as the high volume default, a frontier model for the hardest jobs, and an open weight option where control or data residency matters. Even inside 3.8 Flash, the low, medium and high reasoning settings give you three different cost and quality profiles.
What Should a Small Business Do About This?
You do not need a gateway or a telemetry stack. The principle still applies at your scale.
Keep your notes, prompts, research and source documents in exportable formats instead of letting the only good copy live inside one AI tool’s workspace. Separate the tool that holds your information from the model that reasons over it.
Pay for one cheap fast default and keep access to at least one strong alternative. Fable 5.1 landed on September 1 and Gemini 3.8 Flash on September 2. Last month’s favorite stops being the obvious pick that fast.
Then spend your real effort on the part that compounds. Write down what you know, document your processes, and build the knowledge base your business has always needed. Those assets get more valuable with every model release. Your choice of model gets less permanent with every one.
If you want help building the knowledge base and content system that makes your business the one AI recommends, request a free digital marketing audit and I will show you where you stand today.