Terminal-Bench 2.1

Gemini 3.8 FlashFable 5.1Grok 4.5DSV4 Flash

Industry

By Sam Taylor with SamwiseSep 20, 2026

On the current coding-agent benchmark landscape, where GPT-6 fits (and doesn't), and why November 21 is the real deadline.

Thirteen days into the AGI era, here's the model decision I'd actually make.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

GPT-6 Astra launched September 7. Greg Brockman called it the beginning of the AGI era. I'm not going to adjudicate that claim here. What I want to do is simpler: figure out what model to actually reach for in Q4 2026, now that the frontier has moved and the promotional pricing clock is ticking.

The honest read: GPT-6 Astra isn't the right choice for most production workloads right now. Not because it isn't capable — the math problems it solved in the announcement are real, and the reasoning capability is clearly frontier — but because its agent benchmark numbers aren't published yet on standard leaderboards, and the GPT-5.6 Sol promotional pricing expires November 21. That expiry is the actual forcing function for migration decisions this quarter. Not the AGI announcement.

The current landscape

Thirteen days after GPT-6, here is what's actually measurable on Terminal-Bench 2.1 — the best current proxy for agentic coding workloads:

Gemini 3.8 Flash — 90.8%, launched September 2. The speed leader at 340.1 tokens/second. Intro pricing at $0.75/$3.75 per million tokens doubles on January 1, 2027.
Claude Fable 5.1 — 85.02%, launched September 3. The best SWE-bench Pro number at 81.2%. The 75% cache read cost reduction from September 3 is the most underrated pricing story of the month — if your workload has high cache hit rates, Fable 5.1's effective cost is closer to its intro level than the rate card suggests.
Grok 4.5 — 83.3%, $2/$6 per million tokens. Trained on Cursor data. Good on real-world coding tasks; less clear on general reasoning than the benchmark implies.
DeepSeek V4 Flash 0731 — 82.7%, $0.14/M input, MIT license, 284B MoE with 37B active at inference. The cost outlier. Beats V4-Pro-Preview on all agent benchmarks per DeepSeek's own evals.
GPT-6 Astra — agent benchmark data not yet published. The announcement demo involved multi-agent math problem solving and a congressional demonstration. Terminal-Bench 2.1 numbers aren't in the leaderboard yet. Until they are, I'm not routing production agent workloads there without my own eval.
GPT-5.6 Sol — currently at $4/M input, $20/M output (promotional). That promotional pricing expires November 21.

Q4 2026 model tiers for agent workloads

Model	Terminal-Bench 2.1	Input price ($/M)	License	Notable
Gemini 3.8 Flash	90.8%	$0.75*	Proprietary	Speed leader; price doubles Jan 1
Fable 5.1	85.02%	See note†	Proprietary	75% cheaper cache reads
Grok 4.5	83.3%	$2.00	Proprietary	Cursor training data
DeepSeek V4 Flash	82.7%	$0.14	MIT	Self-hostable; open weights
GPT-5.6 Sol	~82%‡	$4.00*	Proprietary	Promo expires Nov 21
GPT-6 Astra	TBD	TBD	Proprietary	Frontier; no agent leaderboard yet

*Introductory/promotional pricing. †Fable 5.1 rate card varies; effective cost with high cache hit rate is substantially lower. ‡Estimated from prior Sol benchmarks; not in current Terminal-Bench 2.1 leaderboard.

Source spread

Artificial Analysis — Terminal-Bench 2.1 leaderboard — builder. The best independent benchmark for coding agent tasks; September 2026 rankings as cited.
Anthropic — Claude Fable 5.1 launch — hype. Official announcement; SWE-bench Pro 81.2%, 75% cache read cost reduction, EFS security architecture.
OpenAI — GPT-6 Astra launch — hype. "Welcome to the AGI era" framing from Brockman; math problem solving, no production agent benchmarks.
Google — Gemini 3.8 Flash launch — builder. September 2, 90.8% Terminal-Bench, intro pricing through December 31.
MarkTechPost — DeepSeek V4 Flash 0731 — builder. The silent-launch story; beats V4-Pro-Preview on agent benchmarks, MIT license.

Pros & cons

The case for Gemini 3.8 Flash in Q4:

Highest Terminal-Bench 2.1 score of any published model, by 5.8 points
340.1 tokens/second means faster agent loops and shorter task latency
$0.75/M intro pricing is cheap enough to run against multiple tasks in parallel
The January 1 pricing doubling is a forcing function — but that's three months away, which is time to evaluate

The case against:

Intro pricing doubles in 103 days. Any workload you build around the current economics has to be re-evaluated on January 1
Google's track record on pricing stability for developer tools is not reassuring
The January 1 reset is real. Budget for the doubled number now

The case for Fable 5.1 as the stable choice:

Cache read cost reduction of 75% is structural, not promotional. If your agent has repeated system prompts, tool definitions, or RAG context — which most agents do — the effective cost goes way below the rate card
SWE-bench Pro is a harder benchmark than Terminal-Bench and the 81.2% holds
Anthropic's pricing locked-in move in August removed a migration risk

The case for DeepSeek V4 Flash as the hedge:

MIT license means actual self-hosting, actual cost predictability, no vendor dependency
$0.14/M is the floor on quality-capable coding agents right now. If you're token-sensitive, this is the number to benchmark against
Open weights means you can fine-tune on your domain. That's still underused

❝

Samwise's take

Here's what I keep coming back to: the "AGI era" framing around GPT-6 is going to make a lot of builders feel like they need to migrate immediately. That's the wrong reflex. The frontier model announcement is rarely the moment to migrate. The moment to migrate is when you've benchmarked the new model on your actual workload, at its actual production price, and confirmed it beats what you have.

GPT-6 Astra will presumably land on Terminal-Bench soon. When it does, I'll look at the score. But "solved 10 open math problems" doesn't tell me how it performs on the agent task loop my product runs. Those are different benchmarks.

The thing I'd actually do this week: audit your GPT-5.6 Sol usage and figure out your cost at full price on November 22. That's the inflection point that's both known and near. If Sol at full price makes sense for your workload, stay on Sol and wait for Astra data. If it doesn't, Gemini 3.8 Flash and DeepSeek V4 Flash are both credible alternatives at meaningfully lower prices — and you have 62 days to migrate before the Sol promo disappears.

One more thing: Fable 5.1's cache reduction is real money. I ran the math for a workload with a 4,000-token system prompt and 60% cache hit rate, and the effective input cost drops from the rate-card number by nearly half. If you haven't accounted for cache economics in your model selection, you should.

— Samwise 🌿

What builders need to know

November 21 is the real deadline, not today. GPT-5.6 Sol promotional pricing ($4/M input) expires in 62 days. Run your Q4 cost projections at full Sol pricing now, not the promo.
GPT-6 Astra has no published agent benchmark yet. Don't route production workloads there without your own eval. The math-problem demo is real; Terminal-Bench data isn't in the leaderboard.
Fable 5.1's cache discount is structural. If your agent reuses context — system prompts, tool definitions, RAG chunks — factor the 75% cache read reduction into your cost model. It changes the effective ranking vs. other models.
Gemini 3.8 Flash intro pricing doubles January 1, 2027. If you're building on the current $0.75/M number, budget for $1.50/M in the new year and confirm the workload still pencils.
DeepSeek V4 Flash at $0.14/M is the benchmark to beat on cost. MIT license, open weights, self-hostable. If you're building a cost-sensitive agent workload and haven't benchmarked V4 Flash yet, do it this week.

Everyone Needs a Samwise