Verified as of August 4, 2026
Claude Fable 5 vs Gemini 3.1 Pro
The two best-positioned frontier models of 2026 take opposite trade-offs: Fable 5 leads the agentic-coding benchmarks at a premium price, Gemini 3.1 Pro undercuts it 4-5x and lives inside the Google stack. Here is the verdict by task, with every number sourced and dated.
TL;DR — who should pick which
By task: for agentic coding and long autonomous runs, Fable 5 — it holds the highest published scores (80.3% SWE-Bench Pro vs 54.2%) and was built for them. For multimodal, images, video and Google-stack integration, Gemini 3.1 Pro. For long context, both have 1M windows — Gemini is the cheaper one. For cost-sensitive high volume, Gemini is roughly 4-5x cheaper on list price and has a real free tier in the Gemini app. There is no one right answer; route by task. Full details below.
Last verified
All prices, specs and benchmark figures on this page were checked against the sources cited on August 4, 2026. Gemini 3.1 Pro is a preview model and Google revises its pricing; where a figure could not be confirmed it is marked "not confirmed" with the date.
Specs & pricing side by side
API list prices per million tokens, USD. Fable 5 figures from Anthropic's pricing docs (verified July 27, 2026); Gemini figures from Google's API pricing (verified August 4, 2026). Gemini's rates apply to prompts up to 200K tokens; longer prompts are billed at higher tiers.
| Claude Fable 5 | Gemini 3.1 Pro | |
|---|---|---|
| Release | 2026-06-09 | 2026-02-19 |
| Model ID | claude-fable-5 | gemini-3.1-pro-preview |
| Status | GA | Preview |
| Input (per M tokens) | $10.00 | $2.00 |
| Output (per M tokens) | $50.00 | $12.00 |
| Cached input | $1.00 | $0.20 |
| Batch (50% off) | $5.00 / $25.00 | $1.00 / $6.00 |
| Context window | 1M | 1M |
| Max output | 128K | 64K |
| Multimodal input | Image, PDF | Image, video, audio |
| Data retention | 30 days, mandatory | None |
| Consumer plans | Max / Team Premium | Google AI Pro $19.99/mo |
| Open weights | No | No |
Gemini 3.1 Pro is a preview model; the $4/$18 rate applies to prompts above 200K tokens (pricing per Google's API docs, August 4, 2026). Max-output figure from Google's model reference (65,536 = 64K). Fable 5 max output is 128K per Anthropic's docs.
Benchmark matrix — with sources
Every score is labeled with where it came from. Anthropic and Google both report their own models' results; independent tests are marked as such. Treat any lab-published number as directional.
SWE-Bench Pro — agentic coding
Fable 5: 80.3% · Gemini 3.1 Pro: 54.2% (Google has not published a 3.1 Pro score; the figure is for the older Gemini 3 Pro preview). Source: Anthropic's June 9, 2026 launch announcement for both sides — vendor self-reported, settings differ.
FrontierCode (Diamond) — hard coding
Fable 5: 29.3% · Gemini 3.1 Pro: no published score. Source: Anthropic, June 9, 2026 — vendor self-reported; Google has not published a comparable FrontierCode result, so this is absence of evidence, not a Gemini failure.
GDPval-AA — agentic knowledge work
Fable 5: 1932 · Gemini 3.1 Pro: 1316 (Artificial Analysis measurement, Feb 2026 — independent). The 3.1 Pro figure improved ~100 points over Gemini 3 Pro preview but still trails Fable 5 and Opus 5.
GDP.pdf — dense documents, vision
Fable 5: 29.8% · Gemini 3.1 Pro: 16.7%. Source: Anthropic's June 9 announcement (vendor self-reported); the Gemini figure is for the older 3 Pro preview. Independent tests of Gemini 3.1 Pro's document reasoning show it stronger than 16.7% suggests.
OSWorld-Verified — computer use
Fable 5: 85.0% · Gemini 3.1 Pro: 76.2%. Source: Anthropic's June 9 announcement — vendor self-reported, Gemini's number from the 3 Pro preview. Independent computer-use tests of Gemini 3.1 Pro are not yet published.
Multimodal & reasoning (MMMU-Pro)
Gemini 3.1 Pro ranked #1 on MMMU-Pro (Artificial Analysis Intelligence Index, Feb 2026 — independent). Fable 5's multimodal scores are published but not directly comparable. For images, video and audio input, Gemini is the stronger multimodal stack.
AA-Omniscience — hallucination rate
Gemini 3.1 Pro cut its hallucination rate from 88% to 50% (Artificial Analysis, Feb 2026 — independent; accuracy 53%). No directly comparable published figure exists for Fable 5.
Sources: Anthropic "Claude Fable 5" announcement (June 9, 2026); Google Gemini 3.1 Pro launch (Feb 19, 2026); Artificial Analysis Intelligence Index (Feb 2026); Google model reference (Aug 4, 2026). Independent = run by Artificial Analysis or another third party, not the vendor.
Fable 5's unique angle: the safeguards you don't see
Three things that separate Fable 5 from Gemini 3.1 Pro that no benchmark shows — and one export-control caveat that applies to Fable 5 but not Gemini.
Safeguard & Mythos routing
Fable 5 runs safety classifiers that reroute flagged conversations to Opus 4.8 — under 5% of conversations on average (Anthropic). Unrestricted requests go to the same model released as Mythos 5, which is only available to vetted organizations. Gemini has no such dual-release split.
30-day data retention
Fable 5 has a mandatory 30-day retention window — zero-data-retention agreements do not apply. Gemini 3.1 Pro has no equivalent retention requirement; Google's data-governance options differ and are governed by its own terms.
Export controls
Fable 5 was suspended worldwide June 12 – July 1, 2026 under a US export-control directive, then restored. A repeat suspension is possible in principle. Gemini 3.1 Pro has no comparable history. If continuous availability is your hard requirement, this asymmetry is a real risk factor for Fable 5.
Decision matrix — if you're…
The honest answer is almost always "both, routed by task". Here is the per-scenario call.
You're building coding agents
Fable 5. The published gap is 26 points on SWE-Bench Pro, and multi-hour agentic runs are exactly what Fable 5 was built for. If budget forces a choice, run Gemini for routine steps and escalate the hard 10-20% to Fable 5.
You need images, video, or audio
Gemini 3.1 Pro. It ranked #1 on MMMU-Pro and its native multimodal stack is broader. Fable 5 accepts images too, but Gemini is the stronger default for media-rich workflows.
You're working with long context
Both handle 1M tokens. Gemini is far cheaper for context-heavy traffic, but watch the tiered pricing — prompts above 200K tokens bill at $4/$18. Fable 5's flat $10/$50 is simpler to predict.
You're cost-sensitive or high-volume
Gemini 3.1 Pro, clearly — roughly 4-5x cheaper on list price, plus a free tier in the Gemini app for consumer use and $19.99/month Google AI Pro for full access. Fable 5 on Claude Max (included in up to 50% of weekly limits) only makes sense when you're already on a Claude plan.
You're in a regulated industry
Depends on the constraint. If you need unrestricted capability in sensitive domains, only Mythos 5 (vetted access) provides it — Gemini has no equivalent program. If you need the opposite — no data retention obligations — Fable 5's mandatory 30-day retention is a problem and Gemini wins.
Cost examples — run the numbers
Representative workloads priced at list rates per million tokens. Fable 5 assumes high effort, no cache, no fallback; Gemini assumes standard mode, no cache, prompts under 200K tokens. Your real bills will differ — run them through our cost calculator.
Workload A (coding agent, 1M input / 200K output): Fable 5 ≈ $20.00, Gemini 3.1 Pro ≈ $4.40 — Fable 5 costs ~4.5x. Workload B (long-context RAG, 10M input / 100K output, cached): Fable 5 ≈ $20.00, Gemini ≈ $2.75 — ~7x gap, narrowing to ~3x at 80% cache hit. Workload C (multimodal, 2M input / 100K output): Fable 5 ≈ $25.00, Gemini ≈ $5.20. On any workload both models can do, the price gap is the story.
The benchmark lead does not survive the price gap on work both models can do. The rule of thumb: if you can accept Gemini's quality bar, a 4-5x cheaper model wins on volume; use Fable 5 where cost-per-solved-task matters more than cost-per-token — hard coding, long agent runs, retry-sensitive workloads.
Still deciding?
Run your actual workload through the cost calculator, or see how Fable 5 compares against Anthropic's own lineup.