Frontier Report · Plate No. I · June 2026

Comparing

Three labs. One frontier.

Claude Fable 5, GPT-5.5 and Gemini 3.1 Pro — what actually changed, who wins what, and which one you should reach for.

Contenders3

Benchmarks6

Best modelIt depends

Scroll to examine

01 · The Contenders

Same race, three very different bets.

Each lab optimized for something different this cycle — which is why "which is best" is the wrong first question.

fig. I

Anthropic

Claude Fable 5

Released June 9, 2026

The new apex. First Mythos-class model — a tier above Opus — built for long-horizon agentic work: repo-scale coding, computer use, tasks that unfold over hours, not turns.

Price / M tok$10 in · $50 out

Context1M+ tokens

Signature winAgentic coding

fig. II

OpenAI

GPT-5.5

Released April 2026

The ecosystem heavyweight. Mid-pack price, the strongest published computer-use and terminal scores, and the best proven long-context retrieval at a full million tokens.

Price / M tok$5 in · $30 out

Context1M · strong MRCR

Signature winComputer use

fig. III

Google DeepMind

Gemini 3.1 Pro

Early 2026

The volume king. Cheapest frontier tokens by far, the broadest native multimodality — video, audio, images — and elite science reasoning. Built to be everywhere.

Price / M tok$2 in · $12 out

ContextLarge · 64K out cap

Signature winCost + multimodal

No model wins every row.
The frontier split three ways —
coding, computers, and cost.

Fable 5 took agentic coding by the widest margin in frontier history. GPT-5.5 kept computer use and million-token recall. Gemini undercut them both at a fifth of the price. The mature answer in 2026 is a router, not a religion.

0

Points of Fable 5's lead over GPT-5.5 on SWE-Bench Pro — larger than the gap between GPT and Gemini.

0

The gap on Cognition's FrontierCode — 29.3% vs 5.7% on production-grade engineering tasks.

0

Fable 5's independent hallucination rate — against 50% for Gemini and 85% for GPT-5.5. Lower is better.

02 · The Numbers

Benchmarks, read honestly.

A mix of vendor-reported and independent figures. Cross-vendor gaps of 1–3 points are noise — the gaps below mostly aren't.

tab. i

SWE-Bench Pro

Real GitHub engineering tasks — the closest public proxy for "can it do a software engineer's job".

Higher is better · vendor-reported, Jun 2026

Fable 5

0

Opus 4.8

0

GPT-5.5

0

Gemini 3.1

0

tab. ii

FrontierCode Diamond

Cognition's brutal production-grade coding eval. Nobody aces it — but the spread is the most lopsided result of the year.

Higher is better · independent (Cognition)

Fable 5

0

Opus 4.8

0

GPT-5.5

0

Gemini 3.1

not published

tab. iii

GPQA Diamond

Graduate-level science questions. A genuine near-tie at the top — and Gemini gets there at a fifth of the price.

Higher is better · mixed sources

GPT-5.5

0

Gemini 3.1

0

Fable 5

0

tab. iv

Hallucination rate

AA-Omniscience, fully independent. The most underrated row on this page — when a wrong answer is expensive, this decides.

Lower is better · independent (Artificial Analysis)

Flipped scale — the short bar wins.

Fable 5

0

Gemini 3.1

0

GPT-5.5

0

tab. v

Recall at one million tokens

8-needle MRCR at full 1M load. GPT-5.5 holds where others collapse — Anthropic hasn't published a comparable Fable 5 figure.

Higher is better · published MRCR

GPT-5.5

0

Gemini

0

Fable 5

not published

tab. vi

Price per million tokens

Three strategies in one table: Google buys volume, OpenAI holds the middle, Anthropic charges for quality that reduces retries.

List price, short-context tier · Jun 2026

Claude Fable 5

Input$10 /M

Output$50 /M

GPT-5.5

Input$5 /M

Output$30 /M

Gemini 3.1 Pro

Input$2 /M

Output$12 /M

03 · The Specimen Gallery

What's actually new, pinned and labeled.

The headline behind each launch — and the catch each vendor would rather you skim past.

fig. IV

C

Work that unfolds over hours

Mythos class — a new tier above Opus. Fable 5 is the safeguarded public release of Mythos 5.
Stripe's stress test — a 50-million-line codebase migration finished in a day; scoped at two-plus months for a team.
Vision leaps — reads exact figures off charts, rebuilds apps from screenshots, beat Pokémon FireRed on sight alone.
The catch — priciest tokens of the three, and no native video or audio.

Agentic apex

fig. V

G

A million tokens, actually recalled

Long context fixed — recall holds past 256K where GPT-5.4 collapsed; 74% on 8-needle MRCR at 1M.
Computer use — the strongest native score here, plus a standing Terminal-Bench win from May.
Ecosystem gravity — the deepest tooling and the safest "nobody got fired for it" pick.
The catch — worst independent hallucination score of the three. Verify everything.

Computer use

fig. VI

✦

Frontier quality at two dollars

$2 input tokens — a price the others simply don't answer, halved again in batch mode.
True multimodal — native video and audio in, not just images. The others don't play here.
Elite science — 94.3% GPQA Diamond, independently corroborated.
The catch — trails badly on agentic coding, with the smallest output cap (64K).

Volume king

04 · Interactive

What are you building?

Choose a task. We'll route you to the right model — with the receipts.

Select a task above to receive a recommendation.

05 · The Verdict

The answer is a router, not a religion.

Teams shipping in 2026 don't standardize on one model — they route per task. The cheat sheet:

i.Coding agents & long-horizon engineering

ii.Accuracy-critical work — legal, finance, medical drafts

iii.Vision-heavy tasks — charts, screenshots, UI rebuilds

Claude Fable 5 Anthropic

iv.Browser & computer automation

v.Million-token document retrieval

vi.Teams deep in OpenAI tooling

GPT-5.5 OpenAI

vii.High-volume content, classification, summarization

viii.Video & audio understanding

ix.Science Q&A on a budget

Gemini 3.1 Pro Google DeepMind

In 2026 there is no best model —
only the best model for the job.