Which frontier model should you build on right now?

In late June 2026 the realistic frontier shortlist is two models: OpenAI’s GPT-5.5 and Google’s Gemini 3.1 Pro. GPT-5.5 is the stronger engineering agent — coding, terminal work, and genuine long-context recall. Gemini 3.1 Pro is the cheaper, broader pick — reasoning, multimodal, and roughly 2.5x lower cost. Pick by task shape, not leaderboard rank.

The obvious question is where Claude went. Anthropic’s Claude Fable 5 actually tops the coding benchmarks (SWE-Bench Pro 80.3%), so it would normally lead any list like this. But on June 12, 2026, a US export-control directive suspended Fable 5 and Mythos 5 for any foreign national — and because nationality can’t be verified at the API layer, Anthropic disabled both models for all customers worldwide. As of late June 2026 they remain offline with no restoration confirmed (Anthropic’s notice). Practical takeaway: don’t architect around Fable 5 coming back next week. Ship on what’s live.

What’s live is two strong, very different models:

GPT-5.5 — OpenAI, released April 23, 2026, powers ChatGPT and Codex (OpenAI’s agentic coding tool). Intro post.
Gemini 3.1 Pro — Google, preview released February 19, 2026, available across the Gemini API, AI Studio, Gemini CLI, Vertex AI, Gemini Enterprise, the Gemini app, and NotebookLM. Announcement.

The thesis, stated plainly so you can stop reading if you just want the answer: pick by task shape, not by leaderboard rank. GPT-5.5 is the stronger engineering agent. Gemini 3.1 Pro is the cheaper, broader value-plus-reasoning-plus-multimodal pick. Neither is “better” in the abstract. The rest of this post is about matching the model to the workload in front of you. If you’re still deciding whether you even need a frontier model for your use case, read AI agents vs automation tools first.

Where does GPT-5.5 genuinely win?

GPT-5.5 is built like an engineering agent, and the head-to-head numbers back that up. These are reported — benchmark comparisons report them, and they’re well-aligned across sources — so I’ll attribute rather than state them as gospel.

The widest gap is on real terminal work. Benchmark comparisons report Terminal-Bench 2.0 at GPT-5.5 82.7% vs Gemini 68.5% — a +14.2-point lead. If your agent lives in a CLI, runs commands, reads output, and iterates, that gap is the difference between “mostly works” and “babysits itself.”

On multi-file code resolution, two separate benchmarks both lean GPT-5.5. And they are two different benchmarks — don’t conflate them:

SWE-Bench Pro (the harder one): GPT-5.5 58.6% vs Gemini 54.2% (+4.4).
SWE-Bench Verified (the easier, more saturated one): GPT-5.5 82.6% vs Gemini 80.6% (+2.0).

The most underrated edge is long-context recall, which I’ll dig into properly later. Reported MRCR v2 retrieval at the full 1M context is GPT-5.5 74.0% vs Gemini 26.3% — a 47.7-point gap. Both advertise a million tokens; only GPT-5.5 reliably reads across all of it.

Round it out with tool orchestration (reported MCP Atlas 75.3%) and two verified facts that matter operationally: GPT-5.5’s max output is 128K tokens (double Gemini’s 64K), and it uses roughly 40% fewer output tokens than GPT-5.4 on the same Codex tasks. That token efficiency quietly takes some sting out of its higher price.

OPINION: this is the engineering agent. Coding agents, multi-file refactors, terminal workflows, and long-context tooling belong here. If that’s your product, the premium is usually worth it.

Where does Gemini 3.1 Pro genuinely win?

Now the other side, and it’s a strong one.

Start with the fact that decides a lot of architectures: Gemini 3.1 Pro is roughly 2.5x cheaper across every pricing tier. Verified pricing is $2 input / $12 output per 1M tokens (≤200K context), stepping to $4 / $18 above 200K, with batch around $1 / $6. Compare GPT-5.5 Standard at $5 / $30. That’s not a rounding difference; at volume it reshapes your unit economics.

On reasoning and science, Gemini is genuinely excellent. Verified benchmarks: ARC-AGI-2 at 77.1% (more than double Gemini 3 Pro), GPQA Diamond at 94.3% (a record), and it topped 13 of 16 tracked benchmarks at launch. Head-to-head, reported GPQA Diamond is a near-tie (Gemini 94.3% vs GPT-5.5 93.6%). Competitive programming is a real Gemini strength — reported LiveCodeBench Pro around 2,887 Elo (GPT-5.5’s figure isn’t published). And web browsing reported BrowseComp 85.9%.

Then the modality gap: Gemini 3.1 Pro is natively multimodal — text, image, audio, and video input. Native audio and video input is credited only to Gemini here. If your product takes a phone call, a meeting recording, or a video clip and reasons over it, that’s not a benchmark nicety; it’s table stakes, and Gemini is the model that covers it.

OPINION: this is the value-plus-reasoning-plus-multimodal pick. High-volume workloads, cost-sensitive products, research and reasoning tasks, competitive-programming-flavored problems, and anything touching audio or video belong here.

Build on GPT-5.5 when… / Build on Gemini 3.1 Pro when…

Build on GPT-5.5

Coding agents and multi-file resolution (SWE-Bench Pro 58.6% vs 54.2%)
Terminal/CLI workflows (Terminal-Bench 82.7% vs 68.5%)
Real 1M-token recall (MRCR v2 74.0% vs 26.3%)
Tool orchestration (MCP Atlas 75.3%)
Bigger output ceiling (128K) and token efficiency

Build on Gemini 3.1 Pro

High-volume / cost-sensitive workloads (~2.5x cheaper)
Scientific reasoning (GPQA 94.3%, ARC-AGI-2 77.1%)
Competitive programming (LiveCodeBench ~2,887 Elo)
Web browsing (BrowseComp 85.9%)
Native multimodal — audio AND video input

The win-split, by task shape. Coding numbers are reported by benchmark comparisons; pricing and modality are verified from the primary sources.

How much does the cost gap actually matter?

Capability decides which model can do the job. Cost decides whether you can afford to run it ten million times a month.

The output-token price is where it bites. Verified, per 1M tokens: GPT-5.5 is $30 output / $5 input; Gemini 3.1 Pro is $12 output / $2 input (Standard, ≤200K). Output dominates most production bills, so that $30-vs-$12 line is the one to watch.

Make it concrete. At 100M output tokens per month — a normal number for a busy product, not a hyperscaler — verified math lands at roughly Gemini $1,200 vs GPT-5.5 $3,000. That’s about $1,800/month of difference on a single workload.

The ~2.5x cost gap, made visceral

Output $/1M — GPT-5.5$30

Output $/1M — Gemini 3.1 Pro$12

100M out/mo — GPT-5.5~$3,000

100M out/mo — Gemini 3.1 Pro~$1,200

Verified pricing and the verified 100M-output-tokens/month worked example. Output tokens dominate most real bills.

Two honest caveats. OPINION: GPT-5.5’s ~40% output-token efficiency on Codex-style tasks narrows the real bill somewhat — you’re billed per token, and it emits fewer of them — so don’t read the headline price as the final invoice. But even after that, Gemini still wins on raw price.

The decision framing I use: at low volume, the price gap is noise — pick on capability and move on. At high volume, the gap compounds every month — make Gemini the default and only reach for GPT-5.5 on the specific tasks that need its engineering or recall edge. If you want the full cost-per-task model behind this, I walk through it in how much it costs to add AI to your app.

Which model do I build this on? A quick decision tree

Here’s the fast path, keyed to task shape. Read top to bottom and stop at the first match.

Which model do I build this on?

Coding agent, terminal/CLI, or multi-file resolution?→ GPT-5.5. The widest, most reliable edges live here.
Need genuine recall across very long context?→ GPT-5.5. Don't assume 1M = 1M elsewhere.
Audio or video input, or heavy multimodal?→ Gemini 3.1 Pro. Native audio + video in.
High-volume / cost-sensitive, or reasoning/competitive-programming heavy?→ Gemini 3.1 Pro. ~2.5x cheaper and strong at reasoning.
Genuinely unsure, or prototyping at low volume?→ Start on Gemini for cost; escalate specific tasks to GPT-5.5 when you hit a coding or recall ceiling.

Match the task shape to the model. Start cheap, escalate on a real ceiling.

That last branch is how I’d default a new project: start cheap, escalate deliberately. You don’t owe the expensive model anything until a task earns it. The plumbing for that — routing, fallbacks, and cost caps — is its own discipline, which I cover in running LLMs in production.

The long-context trap: don’t assume 1M equals 1M

If you take one thing from this post, take this.

Both models advertise a 1,000,000-token context window (GPT-5.5 narrows to 400K inside Codex). On the spec sheet they look identical. They are not.

Reported MRCR v2 recall at the full 1M context is GPT-5.5 74.0% vs Gemini 26.3%. A window you can fill is not a window the model reliably reads. You can stuff a million tokens into either one; only GPT-5.5 will dependably find the fact you buried at position 800,000.

The practical implication is sharp. If your product depends on retrieving specifics from a huge context — long call transcripts, a sprawling codebase, a fat legal contract, a months-long support thread — GPT-5.5 is the safer engine even though it costs more. The advertised number is marketing; the recall number is the one your users feel.

Match the spend to the task

The tradeoffGemini ~2.5x cheaper vs GPT-5.5's engineering + recall edge

MRCR v2 recall @ full 1MGPT-5.5 74.0% vs Gemini 26.3%

The warningAdvertised 1M window is not reliable recall

The ruleDon't overpay by default; don't under-buy recall when the task needs it

The core tradeoff in one figure: pay less, or pay for recall when you actually need it.

OPINION, the core tradeoff in one line: pay ~2.5x less with Gemini, or pay for GPT-5.5’s engineering and recall edge when the task actually needs it. Match the spend to the task; don’t overpay by default, and don’t under-buy recall on the one feature that depends on it.

FAQ

Is GPT-5.5 just better at coding? At agentic/terminal and multi-file resolution, yes — benchmark comparisons report Terminal-Bench +14.2, SWE-Bench Pro +4.4, and SWE-Bench Verified +2.0, all toward GPT-5.5. But it’s not a clean sweep: Gemini leads competitive programming with a reported LiveCodeBench Pro of ~2,887 Elo. “Better at coding” depends on which coding.

They both say 1M context — same thing? No. Reported MRCR v2 recall at the full 1M is 74.0% (GPT-5.5) vs 26.3% (Gemini). Same advertised window, very different ability to actually recall across it.

How much cheaper is Gemini, really? Verified, ~2.5x across every tier. Concretely, at 100M output tokens/month it’s roughly $1,200 vs $3,000.

Can GPT-5.5 take audio or video input? Native audio and video input is a Gemini 3.1 Pro capability — it’s natively multimodal across text, image, audio, and video. For those modalities, Gemini is the pick.

Should I wait for Claude Fable 5? No. It’s been offline for all customers since June 12, 2026 with no restoration confirmed. Build on what’s live and revisit if it ever comes back.

My pick

OPINION, first person: my default is Gemini 3.1 Pro — for cost and for breadth across reasoning and multimodal. Then I reach for GPT-5.5 the moment a task turns into a coding agent, a terminal workflow, or a true long-context retrieval problem. That’s not loyalty to either lab; it’s just where each one earns its keep.

The rule holds: pick by task shape, not leaderboard rank — and don’t wait on Fable 5 to make the decision for you. Neither GPT-5.5 nor Gemini 3.1 Pro is strictly “better.” The right answer is whichever matches the workload in front of you, and most real products end up using both. If you want more of how I wire these into shipping products, the rest of my AI engineering guides go deeper. Primary sources once more: OpenAI on GPT-5.5 and Google on Gemini 3.1 Pro.

GPT-5.5 vs Gemini 3.1 Pro: Which Should You Actually Build On?