Claude Mythos 5 vs GPT-5.6 vs Gemini 3.2 vs Qwen 3.7 vs DeepSeek V4.1: The June 2026 Frontier Refresh

At a glance

Five flagships, five different axes of attack. The table below sets the basics; the bar chart and feature matrix that follow show where each model actually leans hardest.

Snapshot as of late June 2026. Pricing tiers shift weekly across providers; context length runs from 200K (Claude, GPT) through 1M (Gemini) to 256K natively for Qwen and 128K for DeepSeek — figures shift between API tiers and weight releases, so confirm against the active model card before committing.

Release timeline

The convergence is not coincidence. Anthropic and OpenAI run roughly quarterly cadences and both had Q2 2026 slots reserved for their respective flagship updates. Google's Gemini cycle had been telegraphed at I/O. What pulled the Chinese cluster into the same window was specifically DeepSeek V4's April 2026 pricing, which compressed the per-token economics for reasoning workloads to a level the other Chinese labs could not ignore. Qwen 3.7 and the V4.1 refresh that immediately followed both ship inside the same fortnight as the Western flagships — that is a deliberate response, not a calendar accident.

The interesting structural fact is that no two of the five models are racing on the same axis. Anthropic and OpenAI are competing for the agentic crown (one leading on coding, the other on autonomous computer use). Google is competing on multimodal price/quality. The Chinese cluster is competing on per-token cost and weights availability. The two-week window is the moment all five bets became legible at once.

Claude Mythos 5 GA (+ Fable 5 preview)

Mythos 5 went GA in mid-June 2026 as the production sibling of a two-line strategy: Mythos for production, Fable for research preview. The split is procurement-friendly by design. Mythos carries the safety-evaluation paperwork an enterprise buyer expects — vulnerability-aware reasoning evals, model-card disclosure, the usual SOC-shaped diligence — and ships with the same stable API surface as the Claude 4 family it replaces. Fable 5 ships in preview alongside it, on the same model architecture but with research-grade post-training extensions (longer thinking traces, experimental tool patterns, sharper exploratory reasoning). The names matter: nobody has to ask whether a model is procurement-ready, because the line tells you.

On the snapshots that are currently visible, Mythos 5 leads on the two axes Anthropic has been winning all year — coding and agentic reasoning. On SWE-bench Verified it lands at the top of this five-model set; on long-horizon coding tasks the lead is wider than the benchmark gap suggests. The computer-use story carries over from the Vercept integration covered in the previous post on computer-use offerings, which means Mythos 5 inherits the OSWorld lead rather than re-litigating it.

Where Mythos 5 does not lead is price. At roughly $15 per million output tokens it is the most expensive of the five by a comfortable margin, and the gap to the Chinese cluster is an order of magnitude. The pitch is unchanged from the Claude 4 era: if your unit economics survive the per-token cost, the quality and the procurement story are the cleanest in the set. If they do not, you read on. For teams using Anthropic primarily as the orchestrator across other components, see agent frameworks for the integration shapes that make the cost math work.

GPT-5.6

GPT-5.6 is OpenAI on schedule. It is not a discontinuous jump — it is the on-cadence point release that consolidates everything OpenAI shipped between GPT-5 and now into one stable surface. Reasoning quality nudges up; tool-use trajectories get tighter; the model card lists the usual incremental wins on math, code, and long-context recall. The headline story is not the eval numbers. It is the depth of the integration with the OpenAI agent stack.

Specifically, GPT-5.6 is the first OpenAI flagship that ships with Codex Background CU and Operator as first-class targets rather than as adjacent products. The same model drives the cloud Chromium in Operator, the long-running development tasks in Codex Background CU, and the in-conversation "Agent" surface in ChatGPT — without per-surface fine-tunes that drift apart between releases. That unification is the autonomy lead: when the agent has to cross from one tool to another inside a task, the model does not change underneath it. For the architectural shape of those surfaces and how they compare to the other computer-use bets, see Post 1, Claude Computer Use vs Codex CU vs Operator vs Gemini.

At roughly $10 per million output tokens GPT-5.6 prices below Mythos 5 but well above Gemini and the Chinese cluster. The position is coherent: if you are buying the integrated autonomy story, you accept the price; if you are buying a model to call from your own orchestrator, the per-token line is the one to compare. Headline coding scores trail Mythos 5 by a few points on the visible benchmarks — small enough that a release or two from now the order could flip again, and small enough that the autonomy delta matters more than the percentage.

Gemini 3.2

Gemini 3.2 is the mid-cycle point release in Google's 3.x line — not a generational change, but a refresh that sharpens the parts of the model where Google has the structural advantage. Multimodal latency is the headline: video understanding, image reasoning, and screenshot grounding all improved measurably over 3.1, and the round-trip on the 1M-token context window is the lowest of any model that supports that length. For products where the agent has to look at a frame, decide, and act before the user notices a pause, Gemini 3.2 is the one whose architecture is shaped for it.

The Workspace integration is the second piece of the bet. Gmail, Calendar, Docs, Sheets, and Drive all get first-party Gemini surfaces, and the 3.2 update tightens those flows — summarization, draft replies, sheet formula synthesis — without changing the model API. The same model drives the consumer Gemini app, the Workspace surfaces, and the Vertex AI tier for enterprise; the trade is that the deepest integration only pays off if your users are already inside the Google ecosystem.

Pricing at roughly $6 per million output tokens puts Gemini 3.2 in the middle of the five — cheaper than the Western pair, more expensive than the Chinese cluster — and the price/quality position is the one Google has been steadily refining. On general reasoning the model trails Mythos 5 and GPT-5.6 on the current snapshots; on multimodal and on price-per-quality for the workloads it targets, it leads. The 1M-token context is wide enough that long-document RAG starts to look like just "fit it in the window" rather than a retrieval problem — useful, with the usual caveat that long-context recall degrades faster than the headline number suggests.

The Chinese cluster

The two leaders inside the Chinese cluster are Qwen 3.7 and DeepSeek V4.1, and they are the reason the late-June window included five flagships rather than three. Qwen 3.7 extends the Qwen3 family's bet on language coverage and hybrid reasoning with sharper agent tool-use post-training, broader multimodal coverage, and an Apache-2.0 license inherited from the rest of the family. DeepSeek V4.1 follows V4's April pricing salvo with a quality bump — same MLA + sparse-attention architecture from the V3 lineage, refined post-training, and the same per-token economics the V4 line forced into the market.

The cluster behind them sweeps wide. Zhipu's GLM-6 ships an open-weights MoE with hybrid reasoning that competes with Qwen on the multilingual axis. Tencent's Hunyuan Large 3 takes the closed-source Chinese flagship slot with strong CN-domain post-training. Baidu's ERNIE 5.1 lands inside the ERNIE 5 lineage with the same enterprise-Chinese pitch ERNIE has carried for years. ByteDance's Doubao Pro keeps the consumer-scale price floor, optimized for Doubao's deployment context inside the ByteDance product surface. None of these four leads on a frontier benchmark, but together they make the cluster too big to ignore on price and too deep to ignore on weights availability.

The pricing pressure is the story. DeepSeek V4's April 2026 release dropped output token pricing to a level the rest of the Chinese cluster could not match without responding inside weeks — and they did. Qwen 3.7's hosted DashScope tier sits at roughly $2 per million output tokens; DeepSeek V4.1's own API floors at roughly $1.40; the rest of the cluster lands inside that range. Compared to the Western trio at $6, $10, and $15, the cluster is an order of magnitude cheaper for the high-volume reasoning workloads it targets. The compression is not a temporary promotion — it is the price floor the cluster is committed to defending.

Weights availability is the second axis where the cluster pulls ahead. Qwen 3.7 ships under Apache 2.0 with weights on Hugging Face and ModelScope, DeepSeek V4.1 ships under MIT on the same channels, GLM-6 ships open-weights from Zhipu's GitHub. For teams that need to self-host — for data-boundary reasons, for offline deployment, or for the unit-economics math that hosted APIs cannot beat at scale — the Chinese cluster is the only one of the five labs offering a frontier-tier checkpoint to download. The Western trio is closed-API only.

Cross-cutting comparison

Reasoning

On the visible reasoning snapshots — SWE-bench Verified, GPQA, the long-horizon agent traces — Mythos 5 leads narrowly, GPT-5.6 trails by a small margin, and Gemini 3.2 sits a tier below the Western pair on pure reasoning while leading on multimodal-bound reasoning tasks. Qwen 3.7 and DeepSeek V4.1 land in the next tier, with Qwen pulling ahead on multilingual reasoning and DeepSeek pulling ahead on cost-adjusted reasoning (the same answer for a fraction of the spend). The gaps between adjacent ranks are small enough that a release or two from now the order could flip; the gap between the top two and the bottom two is wider but closing each quarter. Read these as snapshots, not as durable orderings — see reading benchmarks for the trap of treating leaderboard order as a stable signal.

Computer use

GPT-5.6 is the autonomous computer-use leader at the API level — Codex Background CU and Operator both ship with the model as a first-class target, and the unified-model story (one checkpoint across all three OpenAI agent surfaces) is the autonomy delta the other four cannot match without a separate fine-tune. Mythos 5 inherits the Claude Computer Use post-Vercept architecture and leads on raw OSWorld score, which means it wins on per-task success when you build the harness yourself but trails on integrated agent products you buy off the shelf. Gemini 3.2 holds the browser-tab-inside-Workspace lane and is the fastest of the three for what it targets. Qwen 3.7 and DeepSeek V4.1 do not ship a first-party computer-use product; the open-weights computer-use stack (browser-use, UI-TARS, CogAgent) runs against them, but the gap to the closed leaders is real and the integration burden lands on you.

Multimodal

Gemini 3.2 leads on multimodal across the board — video, image, audio, screenshot — and the 1M-token context is wide enough that video reasoning starts to look like just a long prompt. Mythos 5 and GPT-5.6 ship strong vision-in capabilities (image reasoning, screenshot grounding for their respective computer-use stacks) without trying to match Gemini's video depth. Qwen 3.7 ships a Qwen3-VL sibling with respectable image reasoning under the same Apache 2.0 license, which makes it the only open-weights flagship in the set with serious vision-in capability. DeepSeek V4.1 stays text-only at the flagship tier; the DeepSeek-VL line is a separate sibling and trails the others on raw vision benchmarks. If your product is multimodal-first, the order is Gemini → Qwen-VL → Mythos/GPT → DeepSeek.

Price/perf

The price chart is the cleanest of the comparisons because the units are objective and the spread is an order of magnitude. Mythos 5 at $15 and GPT-5.6 at $10 are the premium tier, paid for quality and integration; Gemini 3.2 at $6 is the middle bet, paid for multimodal and Workspace fit; Qwen 3.7 at $2 and DeepSeek V4.1 at $1.40 are the cost floor, paid for raw token economics. The crossover point that decides which tier you can afford is workload-dependent — see cost, quality, and latency for the framing — but for high-volume agent loops that replay long prefixes on every tool call, the order-of-magnitude gap to the Chinese cluster moves unit economics in ways quality differences do not. Snapshot prices, snapshot tiers; what is durable is the relative ordering, not the digits.

Weights availability

The Western trio is closed-API only — Mythos 5, GPT-5.6, and Gemini 3.2 do not ship weights, and there is no public roadmap to do so. The Chinese cluster is the inverse: Qwen 3.7 under Apache 2.0, DeepSeek V4.1 under MIT, GLM-6 and the rest of the cluster open under permissive or near-permissive licenses (a 14B Qwen 3.7 dense sibling runs on a single H100; a 4-bit quantization of DeepSeek V4.1 fits on an 8×H200 node). For procurement contexts that require weights — air-gapped deployments, data-boundary regimes that cannot route through a vendor API, regulated industries with hard offline requirements — the choice is structural: the cluster is the only option. The trade is the geopolitical and audit overhead of running a Chinese-lab checkpoint inside a Western enterprise; see open vs closed models for the framing that makes this trade-off explicit.

When to pick which

Further reading

On this wiki:

• Llama 4 vs DeepSeek V3 vs Qwen3 vs Mistral Large 3 — the previous frontier post on open-weights flagships; this one refreshes the picture across both open and closed weights as the cluster expanded into mid-2026.
• Claude Computer Use vs Codex CU vs Operator vs Gemini — the architectural shape of the computer-use surfaces that GPT-5.6's autonomy lead and Mythos 5's OSWorld inheritance both build on.
• Choosing a model — the constraint-first selection guide that turns "best model" into "best for this constraint" and makes the axes-not-leaders framing actionable.
• Reading benchmarks — how to read MMLU / GPQA / SWE-bench numbers without getting played by snapshot ordering that flips every release.
• Open vs closed models — the trade-off framing for the structural split between the Western trio and the Chinese cluster.
• Cost, quality, and latency — the triangle that decides which tier's per-token economics actually matters for your workload.

Project sources:

• Anthropic — News (Claude Mythos 5 GA + Fable 5 preview announcements)
• OpenAI — Blog (GPT-5.6 release notes and Codex/Operator integration updates)
• Google DeepMind — Blog (Gemini 3.2 release post and Workspace integration notes)
• Qwen — Homepage (Qwen 3.7 release, model family, Apache 2.0 license)
• Qwen on GitHub (Qwen-Agent SDK, model cards, deployment guides)
• DeepSeek — Homepage (DeepSeek V4.1 release and API pricing)
• DeepSeek on GitHub (V4 lineage tech reports and inference recipes)
• llm-stats.com — third-party leaderboard cross-checks for the snapshot numbers in this post.