TL;DR
Qwen 3.8 Max is the best open-weights coding model you can try today, but it is not the best coding model — GPT-5.6 Sol and Claude Fable 5 still lead on independently verified benchmarks. Qwen's own release package shows 86.6% on Terminal-Bench 2.1, behind GPT-5.6 Sol at ~88.8% but ahead of Claude Opus 4.8 and Fable 5 at 84.6%. Meanwhile Fable 5 dominates SWE-bench Pro at 80.0% versus Qwen's 67.7%. The real differentiator is price and openness: Qwen 3.8 Max costs roughly $2/$6 per million tokens (international API) and goes open-weights; Fable 5 is $10/$50 and closed. For builders who need frontier capability now, Fable 5 and GPT-5.6 Sol remain the safe pick. For teams optimizing cost-per-token or planning self-hosting, Qwen 3.8 Max and Kimi K3 are the credible alternatives.
Last verified: August 4, 2026. Benchmark scores are vendor-reported unless stated otherwise. Pricing and availability change fast — always check the primary source before budgeting.
At a Glance: The Four Models Compared
| Dimension | Qwen 3.8 Max | Claude Fable 5 | Kimi K3 | GPT-5.6 Sol |
|---|---|---|---|---|
| Maker | Alibaba (Qwen) | Anthropic | Moonshot AI | OpenAI |
| Total params | 2.4T (MoE) | Undisclosed | 2.8T (MoE) | Undisclosed |
| Active params | ~95B per token | Undisclosed | ~50B per token | Undisclosed |
| Context window | 1M tokens | 1M tokens | 1M tokens | Undisclosed |
| Pricing (in/out per 1M) | ~$2 / ~$6 (intl API) | $10 / $50 | From $0.55 / $2.19 (hosted) | $5 / $30 (estimated) |
| Open weights | Announced (not yet shipped) | No | Yes (shipped Jul 26, 2026) | No |
| License | Apache 2.0 (planned) | Proprietary | Open (Modified MIT) | Proprietary |
| Terminal-Bench 2.1 | 86.6% (vendor) | 84.6% (vendor) | 88.3% (per BenchmarkList) | 88.8–91.9% (vendor) |
| SWE-bench Pro | 67.7% (vendor) | 80.0% (vendor) | Not published | 64.6% (vendor) |
| Independent verification | None yet | #3 on Artificial Analysis | Partial | Partial |
Sources: Qwen launch blog (qwen.ai), Anthropic pricing page (anthropic.com), Moonshot/Kimi K3 platform docs (platform.kimi.ai), BenchLM Terminal-Bench 2.1 leaderboard (benchlm.ai), BenchmarkList (benchmarklist.com).
What Is Qwen 3.8 Max and Why Does It Matter?
Qwen 3.8 Max is Alibaba's flagship AI model, released August 3, 2026 as a hosted API on QwenCloud, with open weights promised "the following week." It is a 2.4-trillion-parameter Mixture-of-Experts (MoE) model — meaning only a subset of parameters (approximately 95 billion) activate per token, keeping inference costs down relative to a dense model of the same size. The model supports a 1-million-token context window and multimodal input (text and images). Qwen reports it runs reasoning across 16-day autonomous task windows. The international API is priced at roughly $2 per million input tokens and $6 per million output tokens. Critical caveat: the benchmark numbers below are Qwen's own vendor-reported results, not independently audited. BenchLM and BenchmarkList correctly label these as "provider exact" — meaning the score comes from the vendor's own harness, not a neutral third party.
How Do These Four Models Actually Compare on Coding Benchmarks?
On Terminal-Bench 2.1 — the terminal-based software-agent benchmark — GPT-5.6 Sol leads at approximately 88.8% (Qwen's own table) to 91.9% (BenchmarkList field leader), with Qwen 3.8 Max at 86.6%, Kimi K3 at 88.3%, and Claude Fable 5 at 84.6%. But SWE-bench Pro — which tests real-repository bug fixing — reverses the order: Fable 5 dominates at 80.0%, Qwen 3.8 Max sits at 67.7%, and GPT-5.6 Sol falls to 64.6%. This split is the single most important data point in this comparison, and it reveals why no single benchmark is trustworthy for a coding-model decision.
| Benchmark | Qwen 3.8 Max | Claude Fable 5 | Kimi K3 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 86.6% | 84.6% | 88.3% | 88.8–91.9% |
| SWE-bench Pro | 67.7% | 80.0% | Not published | 64.6% |
| PaperBench | 93.0% | 90.5% | Not published | 90.5% |
| OSWorld-Verified | 86.1% | Not published | Not published | Not published |
| Toolathlon | 72.5% | 77.9% | 73.2% | 74.9% |
| Agents' Last Exam | 27.0% | 40.5% | Not published | 30.6% |
Sources: Qwen3.8-Max release benchmarks (qwen.ai), BenchmarkList cross-model comparison (benchmarklist.com), BenchLM (benchlm.ai), evolink.ai vendor-score analysis (evolink.ai).
What These Scores Actually Mean
Terminal-Bench 2.1 measures terminal-based software-agent tasks — executing commands, managing environments, running builds. Qwen 3.8 Max's 86.6% means: under Qwen's own test harness, it successfully completed 86.6% of a set of professional software-engineering tasks in a terminal environment. BenchLM notes these are "display-only" scores because the harness and agent configuration differ from the weighted Terminal-Bench 2.0 lane, meaning they are not directly comparable to earlier results.
SWE-bench Pro measures whether a model can fix real GitHub bugs — open the repo, find the issue, write the patch, and pass the test suite. Fable 5's 80.0% here means it is genuinely excellent at repository-level work. Qwen's 67.7% — over 12 points behind — tells a different story. The evolink.ai analysis is blunt: "a universal coding lead; Qwen's own table includes stronger competitors" on SWE-bench Pro.
Agents' Last Exam is a harder agentic benchmark. Here Fable 5 dominates at 40.5%, with Qwen at 27.0% — a 13.5-point gap that suggests Fable 5 remains meaningfully stronger at complex multi-step agent tasks.
Which Model Should You Actually Use for Coding Work?
For daily-driver coding and production agent deployments today, Claude Fable 5 remains the safest choice — it has the most mature ecosystem (Claude Code, tooling, guardrails) and the best independently audited coding score on SWE-bench Pro. For teams that need open weights or cost optimization, Qwen 3.8 Max is the credible challenger, and Kimi K3 is the open-weights option you can actually download right now.
The Decision Framework
Pick GPT-5.6 Sol if: you want the highest vendor-reported Terminal-Bench 2.1 score, you already use OpenAI's API stack, and the premium pricing is acceptable for your workload.
Pick Claude Fable 5 if: you do real repository-level work (SWE-bench Pro is the benchmark that maps to this), you are building production agent pipelines, you value the Claude-Code ecosystem, or you need 1M context with proven reliability.
Pick Qwen 3.8 Max if: cost-per-token matters significantly to you, you want to evaluate an open-weights model before committing, your work leans toward terminal/build tasks and research-paper reproduction (PaperBench 93.0%), or you want a non-US-vendor option for regulatory diversity.
Pick Kimi K3 if: you want open weights that are actually downloadable today (weights went live July 26, 2026), you need a 1-million-token context window in an open model, you are doing research or fine-tuning work, or you want the largest open model currently available.
How Much Does Each Model Actually Cost?
At published API rates, Qwen 3.8 Max is roughly 5× cheaper than Claude Fable 5 on input tokens and 8× cheaper on output tokens. Kimi K3 via hosted providers is cheaper still. But there are critical caveats: Qwen 3.8 Max's international API pricing is sourced from Chinese-language coverage and may not reflect final global pricing; Kimi K3's hosted pricing varies by provider; and Fable 5 has prompt-caching (90% off cached inputs) and batch (50% off everything) levers that narrow the gap in practice.
| Cost Dimension | Qwen 3.8 Max | Claude Fable 5 | Kimi K3 (hosted) | GPT-5.6 Sol |
|---|---|---|---|---|
| Input per 1M tokens | ~$2 | $10 | ~$0.55 | ~$5 |
| Output per 1M tokens | ~$6 | $50 | ~$2.19 | ~$30 |
| Prompt cache read | Unknown | $1/1M | Unknown | Unknown |
| Batch discount | Unknown | 50% off | N/A | Unknown |
| Self-host possible? | Yes (when weights ship) | No | Yes (weights live) | No |
Sources: Anthropic Fable 5 pricing (anthropic.com/claude/fable), benchlm.ai API pricing tracker (benchlm.ai/anthropic/api-pricing), Kimi K3 hosted pricing (varies by Together AI / Modal; check together.ai and modal.com).
The Self-Hosting Reality
Kimi K3's open weights are live now — 2.8 trillion parameters, quantization-aware training with MXFP4 weights. But "open weights" does not mean "runs on your laptop." A 2.8T-parameter model is approximately 1.4TB at 4-bit quantization. Even Kimi K3's efficient MoE architecture (only 50B active parameters per token) requires multi-GPU infrastructure to run. The explainx.ai analysis puts it plainly: "Can I run it on a single GPU? No." Qwen 3.8 Max, when its weights ship, will face the same constraint. For most builders, "open weights" means "fine-tunable and private inference via a hosted provider" — not "self-hosted on my desktop."
What About Kimi K3 — the Other 2.8T Open Model?
Kimi K3 from Moonshot AI is the model Qwen 3.8 Max is often compared to and the one that actually has open weights available today. Released July 16, 2026 (API), with weights going live July 26, 2026 — slightly ahead of Moonshot's stated July 27 target. It has 2.8 trillion total parameters (more than Qwen's 2.4T), approximately 50 billion active per token (fewer than Qwen's 95B), and a 1-million-token context window. Its architecture uses "Kimi Delta Attention" (a hybrid linear attention mechanism), Stable LatentMoE routing, and quantization-aware training with MXFP4 weights from the supervised fine-tuning stage — not post-training quantization. Day-0 hosting landed on Together AI and Modal.
On the shared benchmarks BenchmarkList tracks between Qwen 3.8 Max and Kimi K3, the picture is mixed: Qwen wins on JobBench (+2.9 percentile points) but Kimi K3 wins on Terminal-Bench 2.1 (+3.5 percentile points), AutomationBench (+5.0), and Toolathlon (+2.1). This is consistent with the overall theme: the top models trade blows across benchmark categories, and no single model is "best" universally.
Source: Moonshot AI Kimi K3 platform docs and tech blog (platform.kimi.ai, kimi.com/blog/kimi-k3), explainx.ai open-weights coverage (explainx.ai), BenchmarkList cross-comparison (benchmarklist.com).
How Do You Test These Models Yourself Instead of Trusting Vendor Scores?
The only honest way to compare these models for your workload is to run a matched test: identical prompts, identical context, identical acceptance criteria, multiple trials, and separate measurements for quality, latency, token cost, and human-intervention rate. The evolink.ai analysis recommends exactly this, and it is the standard practice in any serious ML evaluation.
A Practical Evaluation Protocol
- Pick 10–20 real tasks from your own codebase. Not benchmark puzzles — your actual bugs, features, and refactors. This is the workload that matters.
- Run each task against each model with identical instructions. Same system prompt, same tools, same file context, same submission format. If a model offers different reasoning effort levels, document which you used.
- Score on a rubric, not on vibes. Did the patch apply cleanly? Did tests pass? Did it need human intervention? How many turns did it take? How many tokens did it consume?
- Measure cost per successful task, not cost per token. A model that costs 5× more but solves 2× more tasks is actually cheaper per success.
- Run at least 3 trials per task per model. Variance is real. A single run tells you nothing.
- Report with confidence intervals, not point estimates. "Fable 5 solved 7/10 vs Qwen 3.8's 6/10" is not statistically meaningful. "Fable 5 solved 7/10 (95% CI: 3.5–10) vs Qwen 3.8's 6/10 (95% CI: 2.7–9.3)" is honest about the uncertainty.
If you want to see how a community-driven comparison works in practice, check out our guide to running a fleet of AI coding agents for free with Orca — the same parallel-worktree pattern lets you bench multiple models against the same task simultaneously.
What This Means for You
If you are a solo builder or small team: Start with the cheapest model that can handle your workload. Qwen 3.8 Max's ~$2/$6 pricing and Kimi K3's hosted rates make experimentation nearly free. When you hit walls — a refactor too complex, a bug too subtle — escalate to Fable 5 or GPT-5.6 Sol for that specific task. Model routing by difficulty is the single biggest cost lever available, and most teams underuse it. Read our budget AI model decision guide for the routing logic.
If you are an agency or team shipping AI products: Do not pick one model. Set up model routing so hard tasks go to Fable 5, routine tasks go to Qwen or Kimi, and you measure the savings. Our agent operating system decision framework walks through the architecture. The teams winning on cost in 2026 are the ones that route intelligently, not the ones that pick one "best" model and overpay for every call.
If you need open weights for compliance, privacy, or fine-tuning: Kimi K3 is your option today — weights are live, the license is permissive, and the architecture is well-documented. Qwen 3.8 Max promises Apache 2.0 weights "the following week" (as of August 3, 2026), which would make it the more commercial-friendly license. When Qwen's weights do ship, read our deep dive on the Qwen 3.8 Max open-weights release for specs and verification status.
If you are at an enterprise with regulatory constraints: Fable 5 offers "US-only inference at 1.1× pricing" for data-residency requirements. Qwen models face ongoing US policy uncertainty around Chinese open-weight model restrictions. Read our coverage of the open-weights policy fight for context. Your procurement team should track both the technology and the policy.
FAQ
Q: Is Qwen 3.8 Max actually better than Claude Fable 5 at coding?
A: On Terminal-Bench 2.1, Qwen reports 86.6% versus Fable 5's 84.6% — a 2-point edge. But on SWE-bench Pro (real repository bugs), Fable 5 leads 80.0% to 67.7% — a 12.3-point gap. On Agents' Last Exam, Fable 5 leads 40.5% to 27.0%. The honest answer: Qwen 3.8 Max is competitive on terminal tasks and research reproduction, but Fable 5 remains stronger on repository-level coding and complex agent work. All Qwen scores are vendor-reported, not independently audited.
Q: Can I run Qwen 3.8 Max or Kimi K3 on my own hardware?
A: Kimi K3's weights are downloadable now, but a 2.8-trillion-parameter model needs approximately 1.4TB of storage at 4-bit quantization and multi-GPU inference infrastructure — not a laptop or a single GPU. Qwen 3.8 Max's weights are not yet released as of August 4, 2026, but will face similar constraints. For most builders, "open weights" means "fine-tunable and available via hosted providers like Together AI or Modal" — not "self-hosted on your desktop."
Q: How much cheaper is Qwen 3.8 Max compared to Claude Fable 5?
A: At published rates, Qwen 3.8 Max costs roughly $2/$6 per million tokens (international API) versus Fable 5's $10/$50 — a 5× difference on input and ~8× on output. However, Fable 5's prompt caching cuts cached input costs by 90% and the Batch API halves everything. In production agentic workloads with caching, the effective gap is smaller. Qwen's international pricing is sourced from Chinese-language coverage and may shift when the global API matures.
Q: What is the difference between Terminal-Bench 2.1 and SWE-bench Pro?
A: Terminal-Bench 2.1 measures terminal-based software-agent tasks — running commands, managing environments, executing builds. SWE-bench Pro measures real-repository bug fixing — opening a GitHub repo, finding the issue, writing a patch, and passing the test suite. The two benchmarks can rank models differently because they test different skills: terminal execution versus repository comprehension. A model strong at Terminal-Bench but weak at SWE-bench is good at "doing things in a shell" but less good at "understanding a real codebase."
Q: Which model should I pick if I want the absolute best coding agent right now?
A: Claude Fable 5, for most production deployments, based on its #3 Artificial Analysis Intelligence Index ranking, its 80.0% SWE-bench Pro score, its mature Claude Code ecosystem, and its production-proven guardrails. GPT-5.6 Sol leads Terminal-Bench 2.1 but has a weaker SWE-bench Pro score (64.6%). Qwen 3.8 Max and Kimi K3 are credible challengers — especially on cost — but lack independent benchmark verification and mature tooling ecosystems. If you want a deeper breakdown of the Qwen 3.8 Max model specifically, read our Qwen 3.8 Max open-weights analysis.
Q: Are the benchmark numbers in this article independently verified?
A: No — with the exception of Fable 5's Artificial Analysis ranking and SWE-bench Verified results, all scores in the Qwen 3.8 Max release package are vendor-reported. BenchLM labels them "provider exact" and "display only." BenchmarkList computes percentile rankings across vendor-reported fields, which is useful for directional comparison but not independent verification. The evolink.ai analysis is explicit: "broad independent replication of the August 3 production model and its new benchmark table is still developing." Treat every number here as "vendor-claimed until independently confirmed."
Discussion
0 comments