Gemini 3.8 Flash: Google's Cheap Model That Matches Opus 5 on Coding
Gemini 3.8 Flash matches Claude Opus 5 on coding benchmarks like Deep SWE at a fraction of the price. Here's how it stacks up.
What is Gemini 3.8 Flash and why does it matter?
Gemini 3.8 Flash is Google’s latest small/fast model in the Gemini family, released just three weeks after Gemini 3.7 Flash and six weeks after the previous flash update. On the Deep SWE benchmark, a long-horizon software engineering test that tracks closely with how coders actually feel about a model, it scored 73.7%, putting it roughly even with Claude Opus 5 (around 74%) and ahead of GPT 5.6 Sonnet (72.7%). The catch, and the reason it matters, is cost: it does this at a small fraction of what Opus 5 or Claude Fable (Opus) charges per task.
TL;DR
- Deep SWE parity with Claude Opus 5 puts Gemini 3.8 Flash at roughly 73.7% to 74%, ahead of GPT 5.6 Sonnet’s 72.7%, on the benchmark most correlated with real-world coding feel.
- Cost per task is the real story: multiple sources put Gemini 3.8 Flash around $2.36 to $2.40 per task on Deep SWE, versus roughly $11 to $12 for Claude Opus 5 and over $20 for the older Claude Fable model.
- Pricing lists at 75 cents per million input tokens and $3.75 per million output tokens, though Google’s own fine print flags these as introductory rates that may rise (to roughly $1.50/$7.50) after a stated cutoff around year-end.
- Benchmark performance is uneven: it tops Humanity’s Last Exam (55.9%) and the Harvey legal benchmark, but lags well behind Opus 5 on GDPval (knowledge work, 1545 vs Opus 5’s 1824), Terminal Bench 4.0 (19.1% vs Opus 5’s over 50%), and OSWorld computer-use tasks (59% vs 75%).
- Harness choice changes outcomes: the same prompt run in the plain Gemini app produces noticeably weaker output than running it through a coding-focused harness like Google’s Antigravity, which supports multi-turn build-verify-update loops.
- A companion cyber-focused variant, Gemini 3.8 Flash Cyber, is restricted to trusted-partner access and scores competitively with specialized cyber models like GPT 5.5 Cyber on the CyberGym benchmark, at a much lower cost than Anthropic’s Fable-class models.
- This is Google’s third flash release in six weeks, a release cadence that signals a deliberate push toward cheap, fast, “good enough” models for production workloads rather than chasing frontier-only benchmarks.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How does Gemini 3.8 Flash compare to Claude Opus 5 and GPT 5.6?
On Deep SWE specifically, the three models land close together: Gemini 3.8 Flash at 73.7-74%, Claude Opus 5 at around 74%, and GPT 5.6 Sonnet at 72.7%. That’s a near-tie among models that occupy very different price tiers. Claude Fable (Anthropic’s larger Opus-class model), which was considered the best model available just weeks before Gemini 3.8 Flash launched, scored around 69-70% on the same benchmark while costing roughly $20+ per task, compared to Gemini 3.8 Flash’s roughly $2.36.
Outside of Deep SWE, the picture is less flattering for Google. On GDPval, a benchmark built by OpenAI to test real-world knowledge work like PDF extraction, data analysis, and presentation creation, Gemini 3.8 Flash scored 1545, well behind Claude Opus 5’s 1824 and GPT 5.6 Sonnet’s 1710. On Terminal Bench 4.0, a newer and harder agentic-terminal-coding benchmark, Gemini 3.8 Flash scored 19.1%, compared to Opus 5’s dominant performance above 50%. On OSWorld, a computer-use benchmark testing a model’s ability to control a browser or desktop, it scored 59% against Opus 5’s 75%.
So the honest summary is: Gemini 3.8 Flash is a coding specialist that punches well above its price class on the specific benchmark most coders care about, but it’s not a general frontier replacement for Opus 5 across knowledge work, terminal agentic tasks, or computer-use tasks.
Why is the price so much lower?
Gemini 3.8 Flash is listed at 75 cents per million input tokens and $3.75 per million output tokens. Compare that to Claude Opus 5 at roughly $5.25 (blended rate cited across sources) or GPT 5.6 Sonnet at around $4.20, and Gemini 3.8 Flash is a fraction of the cost per raw token.
There’s an important caveat buried in Google’s own materials: the 75 cent / $3.75 pricing is flagged as an introductory rate, with a note suggesting it may increase to roughly $1.50 input / $7.50 output after a cutoff around the end of the year. Even at that higher rate, it would still undercut GPT 5.6 Sonnet and Claude Opus 5 substantially.
One wrinkle worth knowing: analysts at Artificial Analysis noted that Gemini 3.8 Flash uses more output tokens per task than its 3.7 Flash predecessor, an increase cited around 30%. That partially offsets the per-token savings, since a cheaper model that needs more tokens to finish a job can end up costing close to what a pricier, more token-efficient model would. The cost-per-task figures on Deep SWE already account for this and still show Gemini 3.8 Flash coming out well ahead on value.
Does the coding harness actually change the output?
Yes, and this is one of the more practical findings from early testing. Running identical prompts through the standard Gemini chat app versus a dedicated coding harness like Google’s Antigravity tool produced meaningfully different results. In the plain app, outputs for complex, multi-step prompts (building a 3D simulation, for example) came back simpler and less polished. In Antigravity, which supports a build-verify-update loop where the model implements a step, checks its own work, then revises, the same prompts produced noticeably richer results: added sound effects, physics interactions, day/night lighting changes, and more layered detail.
This matters for anyone benchmarking or evaluating models informally: the model isn’t the only variable. The scaffolding around it (how many turns it gets, whether it can verify its own output, what tools it has access to) can swing perceived quality significantly. Testing a model only in a basic chat interface may understate its real capability for coding tasks.
Is Gemini 3.8 Flash worth using for coding projects?
For teams optimizing production workloads where cost per task matters as much as raw capability, yes. It sits near the Pareto frontier of cost versus performance on Deep SWE, meaning it’s hard to find a model that’s both smarter and cheaper on that specific axis. For well-defined, scoped coding tasks, agentic loops, and high-volume workloads where paying Opus 5 or Claude Fable prices repeatedly isn’t sustainable, it’s a reasonable default.
For open-ended creative coding, complex knowledge work, or long agentic terminal sessions, the gap to Opus 5 is real. Terminal Bench 4.0 and GDPval scores suggest it still trails frontier models when tasks require sustained multi-step reasoning outside pure software engineering. Informal side-by-side tests building small 3D scenes, simple websites, and game prototypes generally placed its output above older flash models but behind top-tier models like GPT 5.6 Sonnet or Claude Opus 5.1 on visual polish and detail, even when it matched or beat them on raw coding benchmarks.
What is Gemini 3.8 Flash Cyber?
Alongside the general release, Google also introduced Gemini 3.8 Flash Cyber, a variant built specifically for cybersecurity defense work: vulnerability discovery, code auditing, and related tasks. It’s not generally available; access is limited to a trusted-partner program. On the CyberGym benchmark, it scored competitively with specialized security models like GPT 5.5 Cyber, and Google also tested it on an internal benchmark spanning 20 programming languages (beyond CyberGym’s C/C++ focus), where it showed a large jump over the previous flash generation. The pitch is similar to the main model: performance close to expensive frontier options, at a much lower cost.
Frequently Asked Questions
What is Deep SWE and why does it matter for judging coding models?
Deep SWE is a benchmark for long-horizon software engineering tasks. It’s widely regarded as one of the better predictors of how a model actually feels to use for real coding work, more so than broader intelligence benchmarks, because it tests sustained, multi-step engineering rather than isolated problem-solving.
How does Gemini 3.8 Flash’s price compare to Claude Opus 5 in practice?
Per-task cost estimates on Deep SWE put Gemini 3.8 Flash around $2.36 to $2.40 versus roughly $11 to $12 for Claude Opus 5, meaning Gemini can complete similar coding tasks at close to a fifth of the cost, though Opus 5 still leads on several other benchmarks like GDPval and Terminal Bench.
Is the 75 cent / $3.75 pricing permanent?
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
No. Google’s own documentation flags this as an introductory rate that is expected to increase, with figures suggesting roughly $1.50 input and $7.50 output tokens after a stated cutoff near the end of the year, though the company could extend the lower pricing if demand is strong.
Should I test Gemini 3.8 Flash in the Gemini app or a coding harness?
A dedicated coding harness, such as Google’s Antigravity tool, produces noticeably better results for complex or multi-step coding prompts than the standard chat app, since it supports iterative build-and-verify loops the base app doesn’t offer.
Is Gemini 3.8 Flash better than GPT 5.6 Sonnet?
On the Deep SWE coding benchmark, Gemini 3.8 Flash edges out GPT 5.6 Sonnet (73.7% to 74% versus 72.7%). On other benchmarks like GDPval, results are closer or favor Sonnet, so the answer depends heavily on the specific task type.
