Best AI Models August 2026: The US-China Gap Just Tightened

The State of AI Models in August 2026

Keeping up with AI model releases in 2026 feels like drinking from a fire hose. In the last eight weeks alone, Anthropic released four new models, OpenAI cut flagship prices twice, a US export control order took the most capable public model offline worldwide, and a Beijing startup's open weight release moved the Nasdaq. Now, in early August, DeepSeek has quietly demonstrated that a retrained small model can outscore its own larger flagship, without changing a single parameter.

This guide is the current ranking of the 10 best AI models, with verified pricing and benchmarks, and, most importantly, what any of it means for you. We review and update it continuously; this edition was verified against primary sources on August 9, 2026.

---

The 10 Best AI Models Right Now

Ranked by the Artificial Analysis Intelligence Index, the most widely cited composite benchmark in the industry, as tracked through early August 2026. Prices are per million tokens (input/output). For live pricing and side-by-side comparisons, use our AI model comparison tool.

Rank	Model	Creator	AA Index	Price (1M tokens)	Context	Released
1	Claude Opus 5	Anthropic	60.7%	$5.00 in / $25.00 out	1M	Jul 24
2	Claude Fable 5	Anthropic	59.9%	$10.00 in / $50.00 out	1M	Jun 9 (restored Jul 1)
3	GPT-5.6 Sol	OpenAI	58.9%	$5.00 in / $30.00 out	1.05M	Jul 9 (GA)
4	Grok 4.5	SpaceXAI	54.0%	$2.00 in / $6.00 out	500K	Jul 8
5	Claude Sonnet 5	Anthropic	Not yet indexed	$2.00 / $10.00 (through Aug 31)	1M	Jun 30
6	Kimi K3	Moonshot AI	Not yet indexed	Open weight, self-hosted or via API	Unconfirmed	Jul 27
7	GLM-5.2	Zhipu / Z.ai	Not yet indexed	$1.40 in / $4.40 out	1M	Jun 13
8	Gemini 3.6 Flash	Google	Not yet indexed	$1.50 in / $7.50 out	1M	Jul 21
9	Qwen3.8 Max	Alibaba	Not yet indexed	Not yet published	Not yet published	Aug 2
10	DeepSeek V4	DeepSeek	Not yet indexed	$0.14 to $0.435 in / $0.28 to $0.87 out	1M	Apr 24 (Flash refreshed Jul 31)

A note on how to read this table. The Artificial Analysis Index only scores models with enough independent evaluation volume to settle into a stable composite. Kimi K3, GLM-5.2, and the rest of the field below rank 4 are recent enough, and evaluated across different benchmark suites, that they have not yet been indexed the way the closed frontier models have. That does not mean they are weak. It means the industry's standard measuring stick has not fully caught up to them yet, which is itself part of 2026's story.

The DeepSeek row deserves a second look, because it currently spans two different models at two different prices. DeepSeek-V4-Flash-0731, released July 31 at $0.14 input and $0.28 output per million tokens, is the officially refreshed version. DeepSeek-V4-Pro-Preview, still priced at $0.435 input and $0.87 output, remains an unrefreshed preview build from April. That gap is the subject of the next section.

---

August 2026: The DeepSeek Story Nobody Saw Coming

The most interesting AI story of early August is not a new model. It is what happened when DeepSeek retrained an old one.

On July 31, 2026, DeepSeek moved its V4-Flash API out of preview and into official public beta under the build name DeepSeek-V4-Flash-0731. The architecture and parameter count did not change at all: it remains a 284 billion total parameter, 13 billion active Mixture-of-Experts model, identical in scale to the version DeepSeek shipped back in April. DeepSeek's own documentation is explicit that the update "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained."

The result of that retraining alone is striking. On Terminal-Bench 2.1, a benchmark for terminal-based agentic coding, the 0731 build scores 82.7, ahead of DeepSeek's own larger and more expensive V4-Pro-Preview, which scores 72.1 on the same test. On DeepSWE 1.1, the jump is even sharper, from roughly 7.3 in the original Flash preview to 54.4 in the retrained build. DeepSeek published this pattern across nine separate agent and coding benchmarks, and in every case the smaller, cheaper, freshly retrained model outperformed its own bigger sibling.

Pricing for Flash stayed flat at $0.14 input and $0.28 output per million tokens, and the migration was silent: anyone already calling deepseek-v4-flash received the upgraded model automatically with no code changes required.

Meanwhile, DeepSeek's own account has confirmed that the official release of V4-Pro is coming, stating plainly that it is "coming ASAP." As of this writing, that release has no confirmed date, and the V4-Pro API, along with the app and web chat models, remain on the original April 24 preview build. If DeepSeek applies the same re-post-training approach to Pro that it applied to Flash, the current V4-Pro-Preview numbers, including an 80.6 percent score on SWE-bench Verified and a Codeforces rating of 3,206, should be treated as a floor rather than a ceiling for what the eventual official release delivers. For the complete breakdown of what changed, what did not, and what to expect next, see our dedicated coverage of DeepSeek's V4 Pro official release.

The broader takeaway matters beyond DeepSeek's own lineup. Post-training a model this size is dramatically cheaper than pretraining a new one from scratch, and DeepSeek just demonstrated a meaningful capability jump using that lever alone, with no new hardware, no larger model, and no price increase. It is a preview of how open weight labs may keep closing the gap with closed frontier models over the rest of 2026, through better reinforcement learning technique, not just bigger training runs.

---

Other August Releases Worth Knowing About

The pipeline stayed full beyond DeepSeek. Qwen3.8 Max launched August 2, followed by Qwen Image 3.0 and 3.0 Pro on August 5, and Meta's Muse Spark 1.2 on August 6. Kimi K3's benchmark claims from Moonshot AI continue to move through independent verification at Artificial Analysis and LMArena, which is 2026's most consequential ongoing benchmark story, since both evaluators need real evaluation volume before assigning a stable, trustworthy score.

Away from the flagship language models, Microsoft made its first major cybersecurity push since its leadership shake-up, unveiling MAI-Cyber-1-Flash in late July. Paired with GPT-5.4, Microsoft says it outperforms Mythos 5, Gemini 3.5 Flash Cyber, and GPT-5.5 Cyber on CyberGym at half the cost. It will power Project Perception, a vulnerability discovery and patching tool entering early testing in August.

The regulatory backdrop kept building too. Sam Altman spent late July on Capitol Hill previewing OpenAI's next model family, confirmed OpenAI deactivated a model after a Hugging Face breach, endorsed the "Pacing the Frontier" letter, and raised OpenAI's 2026 capital expenditure outlook to $130 to $145 billion. The same week, cybersecurity agencies from the US, Australia, Canada, New Zealand, and the UK jointly published guidance called "Careful Adoption of Agentic AI Services." Governments are no longer just watching AI development. They are actively shaping its release cadence, and that is now a planning input every major lab has to account for.

---

The US-China AI Gap, Honestly Assessed

It is worth resisting both extremes of the current commentary. Kimi K3 did not erase the gap between Chinese and US AI labs. Claude Fable 5 and GPT-5.6 Sol remain ahead of every Chinese model tested on the hardest, most comprehensive benchmarks, and that lead is real, measured in double-digit percentage points on evaluations like SWE-bench Pro.

But the assumption that carried US policy and market pricing through most of 2025 and early 2026, that Chinese labs trailed by six to twelve months with a gap that was stable or widening, no longer holds for the specific categories of coding and general agent capability. GLM-5.2 in June, Kimi K3 in July, and DeepSeek's retrained V4-Flash in early August each demonstrated the same underlying pattern: a fully open, freely downloadable model from a Chinese lab can match or beat the previous generation of US flagship models on real benchmarks, at a fraction of the operating cost, within weeks of a comparable US release rather than months.

What has not changed: the very top of the market, meaning the newest, most expensive, most restricted models like Fable 5, Mythos 5, and GPT-5.6 Sol, remains a US-led category. What has changed: the gap between that top tier and the best available open weight alternative has compressed from a wide margin to single digits on several important benchmarks. For any organization making a build-versus-buy decision on AI infrastructure this month, that compression is the single most consequential fact in this article.

---

What to Actually Use Right Now

Practical guidance based on what is confirmed and priced as of August 9, 2026.

For the best all-around model most teams should run daily: Claude Opus 5 at $5/$25 per million tokens. Near-Fable-5 intelligence, half the price, no mandatory data retention, and it is already the default on Claude Max.

For the single highest ceiling on hard reasoning, legal, and biology work: Claude Fable 5 at $10/$50, if the cost and 30-day retention terms work for your use case.

For agentic coding at the lowest cost per correct answer: GPT-5.6 Sol or Terra, especially after the July 30 price cuts. Sol's Agents' Last Exam lead over every Anthropic model is the most concrete evidence available.

For long-horizon agent reliability on a budget: Grok 4.5 at $2/$6, particularly for teams already inside the Cursor ecosystem.

For self-hosted, open weight deployment with genuine frontier-adjacent capability: Kimi K3, GLM-5.2, or DeepSeek-V4-Flash-0731. Test all three against your specific codebase before committing, since Moonshot's and DeepSeek's benchmark claims are self-reported and independent verification is still accumulating. For hardware sizing, see our Best GPU for Local LLMs and Qwen 3.6 RAM requirements guides.

For everyday text and coding at Sonnet pricing: Claude Sonnet 5 at $2/$10 through August 31, closing in on Opus 4.8's coding performance at a fraction of the cost.

Compare every model side by side on Renovate QR, with full benchmark scores, live pricing, context windows, and deployment notes across every model ranked here.

---

What Comes Next

Gemini 3.5 Pro is the only officially announced model without a release date; Google DeepMind lists it as coming soon. DeepSeek's official V4-Pro release has been confirmed by DeepSeek itself as coming, with no date attached yet. Unannounced but widely expected: GPT-6 (OpenAI has confirmed a new frontier model is in training, and Altman has been previewing the next family on Capitol Hill), Grok 5 (xAI continues training on the Memphis supercomputer, targeting the top spot by Q4), and a sanitized commercial release of Claude Mythos, which analysts expect ahead of Anthropic's anticipated IPO.

Also watch for independent third-party benchmark verification of Kimi K3's claims to firm up through August, whether DeepSeek's official V4-Pro repeats the Flash-0731 retraining jump, and any expansion of Mythos 5 beyond its current restricted list, the clearest signal available on how the export control situation is evolving.

The pattern this year has been consistent: whatever model leads the leaderboard in any given month rarely holds that position for more than a few weeks. Plan your infrastructure accordingly, and keep two or three models on hand rather than betting everything on a single provider.

---

The Bigger Picture

A few themes emerge from 2026's model releases so far.

Performance gaps are narrowing. Opus 5, Fable 5, and GPT-5.6 Sol are all world-class. Choosing between them increasingly comes down to workflow fit, ecosystem, and price, not raw capability. For the reasoning race itself, see our What Is Agentic AI explainer.

Agentic AI is the battleground. Computer use, multi-agent coordination, long-horizon reliability, and autonomous task execution are where differentiation is won and lost. Every major lab is shipping models that act, not just respond.

Open source is serious now, and it just got a new proof point. Llama 4's agentic capabilities, Gemma 4's Apache 2.0 licensing, GLM-5.2's MIT license, Kimi K3's 2.8T open weights, and DeepSeek's retraining-driven jump on V4-Flash together mean local deployment is no longer a compromise. See our Open Source AI vs Closed AI framework for the full decision.

Governments shape release cadence. Export controls, federal benchmarking orders, and restricted access lists are now part of every frontier launch plan. The safest engineering assumption for 2026 and 2027 is that the most capable models will ship late, be gated, or disappear temporarily, so route accordingly.

The best model depends on your problem. The professionals seeing the most value from AI in 2026 are not using one model. They are using different models for different jobs. Our AI Tool Directory lets you filter and compare every major model by benchmark, price, and context window, and our ChatGPT vs Claude comparison walks through the two biggest ecosystems head to head.

---

How We Track and Verify

This ranking is maintained by the Renovate QR editorial team through a systematic monitoring process. We track announcements from OpenAI, Google DeepMind, Anthropic, Meta, xAI, Moonshot AI, Zhipu, DeepSeek, Alibaba, and the rest of the field. We cross-reference claims against independent benchmarks, including ARC-AGI, GPQA Diamond, SWE-Bench, Agents' Last Exam, and the Artificial Analysis Intelligence Index. We verify pricing and availability directly from official vendor pages before any model appears in the tables above, and where a claim comes only from a lab's own self-reported benchmark, such as DeepSeek's V4-Flash-0731 results, we say so explicitly rather than presenting it as independently confirmed.

Review cadence: This article is reviewed on every significant release, pricing change, or access decision. Last verified: August 9, 2026. Major announcements are typically added within 24 hours of confirmation, and the updatedAt date in the page metadata reflects each substantive update.

---

Sources & References

• OpenAI, for GPT-5.6 family, GPT-5.5, and GPT-5.4 official releases and pricing
• Google DeepMind, for Gemini 3.6, 3.5 Flash, and 3.1 Pro
• Anthropic, for Claude Fable 5, Opus 5, Sonnet 5, and Mythos 5
• DeepSeek API documentation, for the V4-Flash-0731 change log and V4-Pro release status
• Moonshot AI, for Kimi K3 release documentation
• Zhipu AI / Z.ai, for GLM-5.2
• Artificial Analysis, for the Intelligence Index and leaderboards
• ARC-AGI, for the abstract reasoning benchmark
• SWE-Bench, for the software engineering benchmark
• Agents' Last Exam, for the long-horizon professional workflows evaluation

---

Last updated: August 9, 2026. Model capabilities and pricing change frequently, so always verify the latest details on official websites. For the most comprehensive and up-to-date comparison, visit our AI Tool Directory or compare models side by side.