Terminal-Bench 2.1 has a leader. Gemini 3.8 Flash hit 90.8% — the highest score from any non-Google model at any price point, and the gap to second place is real: Fable 5.1 at 85.02%, Grok 4.5 at 83.3%, DeepSeek V4 Flash at 82.7%. Five to eight points is a meaningful spread on this benchmark.

The benchmark leader isn't usually the model most builders pick, though. It's a starting point. The actual decision depends on cost structure, license, geopolitical risk, and whether Terminal-Bench 2.1 captures the part of "coding agent performance" that matters for your specific task. Right now, four very different models are within one standard deviation of each other on the leaderboard — which means the selection question is mostly not about the benchmark.

Let me work through each one.

What Terminal-Bench 2.1 actually tests

Before the comparison: knowing what the benchmark measures helps you know when to trust it. Terminal-Bench evaluates models on multi-step coding tasks run in real terminal environments. Write code, run the test suite, read the stderr output, understand what failed, fix it. TB2.1 specifically added tasks requiring multi-file edits, dependency management, and external documentation lookup — things that trip up models optimized for single-file completions.

It's a good benchmark for "can this model operate a coding agent autonomously in a real project." It's less useful for evaluating code review quality, architectural reasoning, or long-horizon planning across multiple sessions. If those are your use cases, TB2.1 is a partial signal. Treat it accordingly.

The four contenders

Gemini 3.8 Flash — 90.8% TB2.1, $0.75/$3.75/M

The benchmark leader. Released September 2, 2026. Knowledge cutoff March 2026. Context window 1M tokens.

The 90.8% is genuine. Gemini Flash models are optimized for speed and tight loops — write-test-debug cycles are exactly the pattern they're built for. The TB2.1 structure rewards that optimization.

The planning item: this pricing doubles January 1, 2027. The current rate is an introductory period. At $0.75/$3.75, Gemini 3.8 Flash is the best performance-per-dollar on the leaderboard by a significant margin. At $1.50/$7.50, the math shifts. If you're building anything that ships before year-end, run the Gemini numbers. If you're planning a 2027 production workload, plan with the doubled rate and verify in November.

Claude Fable 5.1 — 85.02% TB2.1, $10/$50/M, $0.25/M cache reads

The raw benchmark position is second. But the pricing structure is different from every other model in this comparison, and it changes the cost calculation for long-horizon agents.

Cache reads are the dominant cost in a real agentic pipeline. Every turn re-reads the system prompt, tool definitions, accumulated conversation context — all from cache. At Fable 5.1's cache read rate of $0.25/M tokens — 75% below the previous rate — an agent running 1,000 turns against a 50K-token system prompt reads 50M tokens from cache per session. At the old rate: $50 per session. At the new rate: $12.50. The generation cost is identical. The total session cost drops 30–45% depending on your input/output ratio.

There's a break-even point where Fable 5.1 becomes cheaper than Gemini 3.8 Flash per session. That break-even is around 15–20K cached tokens per session, which is well within realistic agentic pipeline sizes. If your system prompt is a 10,000-token repository context plus tool definitions, you're likely past it.

Grok 4.5 — 83.3% Terminal-Bench, $2/$6/M

The one to explain carefully. Released July 14 by SpaceX AI, 1.5T MoE, and trained in part on data from the Cursor acquisition. That last point matters more than it sounds.

Cursor had millions of professional developers working on real production codebases before the SpaceX acquisition closed. That's a different training signal than benchmark-harvested data — actual developer decision-making on actual messy repos, not curated evaluation tasks. Grok 4.5's 83.3% Terminal-Bench score is below both Gemini and Fable, but its real-world behavior on complex, multi-dependency codebases may not reflect that gap accurately. Worth running your own eval if your codebase is on the larger/messier end.

One caveat: some evaluations use Terminal-Bench v1.x for Grok 4.5 rather than TB2.1 specifically. The version isn't always disambiguated in published comparisons. I'd confirm which version you're comparing before treating the 2-point gap as settled.

DeepSeek V4 Flash — 82.7% Terminal-Bench, $0.14/M input

The cost outlier. 284B MoE, MIT license, $0.14/M input, $0.56/M output. That pricing is not a typo.

The MIT license removes commercial restriction questions. Chinese models now account for 30–46% of US enterprise API token usage on OpenRouter, and DeepSeek V4 Flash is the primary driver. The volume-at-scale economics are compelling enough that builders are accepting the geopolitical risk.

That risk is real. Treasury moved on Moonshot AI over the Kimi K3 situation in July. DeepSeek has been treated differently so far — they're not on the Huawei-affiliated list — but if your use case touches defense, healthcare, or government-adjacent anything, model your contingency for disruption.

Source spread

• Google Blog — Introducing Gemini 3.8 Flash [hype] — first-party launch; has the TB2.1 score and the January 2027 pricing note, though the latter is in the FAQ rather than the lede.
• Anthropic — Introducing Claude Fable 5.1 and Mythos 5.1 [hype] — the cache cost reduction is buried in the pricing section; the 85.02% TB2.1 score is in independent evals, not the announcement itself.
• xAI — Grok 4.5 announcement [hype] — the Cursor data flywheel detail is the most substantive claim; benchmark specifics require cross-referencing with independent evaluations.
• Marktechpost — DeepSeek V4 Flash 0731 benchmarks [builder] — most detailed independent benchmark breakdown; has the agentic benchmark comparisons that the other models' announcements don't include.
• Artificial Analysis — coding model benchmark dashboard [skeptic] — the source I use for cross-model comparisons; puts all four on comparable axes and flags when benchmark versions differ.

Pros & cons

What the leaderboard gets right:

• The TB2.1 spread is real signal. An 8-point gap between first and fourth on this benchmark type does translate to observable differences on multi-step coding tasks in controlled settings. It's not noise.
• Gemini 3.8 Flash's lead is earned by something specific — Flash architecture is genuinely better at fast iterative loops, which is what TB2.1 rewards. This will hold until another lab optimizes specifically for that benchmark pattern.
• The four-way competition at this price and capability tier is new. A year ago, this range of options didn't exist. More meaningful competition at the coding-agent level is good for builders.

What the leaderboard misses:

• Terminal-Bench 2.1 is not "what's the best coding agent." It's "what's the best model at Terminal-Bench 2.1." If your agents do architectural review, explain codebases to stakeholders, or handle high-ambiguity tasks, the benchmark correlation to real-world performance is weaker.
• Cost is not in the benchmark. Gemini 3.8 Flash at 90.8% for $0.75/M and DeepSeek V4 Flash at 82.7% for $0.14/M are very different purchasing decisions — a 6-point benchmark gap versus a 5x cost difference — and the benchmark doesn't help you weigh those.
• License and geopolitical risk aren't in the benchmark. MIT vs commercial license, US lab vs Chinese lab, sanctioned-adjacent or not — these are real factors for production deployment that TB2.1 has nothing to say about.

*Grok 4.5 benchmark may reflect Terminal-Bench v1.x in some published evaluations — verify version before direct comparison.

What builders need to know

• Run your own eval before the purchasing decision. Pick 20–30 representative tasks from your actual codebase. All four models offer free tiers or low-cost trial access. An afternoon of testing on your own tasks is worth more than any leaderboard position.
• Model the Gemini January cliff now. If you're building anything on Gemini 3.8 Flash, note the January 1, 2027 pricing change in your system design docs. $0.75/M becoming $1.50/M changes the break-even vs DeepSeek and Fable significantly.
• The Fable 5.1 cache-read math is worth doing. Take your last month of cache read volume from Bedrock or Vertex, multiply by $0.25/M, compare to what you'd pay on any alternative. The 75% cache reduction is automatic on model switch — no integration change required.
• DeepSeek V4 Flash's MIT license matters for some use cases. If you're embedding model outputs in a commercial product and your legal team has been nervous about API ToS restrictions on derived outputs, MIT removes that question. It's not the reason to use it, but it's a real differentiator for certain deployment contexts.
• Don't pick a model for your coding agents without knowing which Terminal-Bench version each score is from. TB2.1 and TB1.x results are published interchangeably in many comparison posts. The version matters for the comparison to be valid.

Further reading

• Google Blog — Introducing Gemini 3.8 Flash — launch details including the January 2027 pricing note
• Anthropic — Introducing Claude Fable 5.1 and Claude Mythos 5.1 — the cache cost reduction announcement
• xAI — Grok 4.5 release — the Cursor training data context
• Marktechpost — DeepSeek V4 Flash 0731 benchmarks — cross-model agentic benchmark comparison
• Artificial Analysis — model benchmark dashboard — the best place for live cross-model comparisons including version-disambiguation