Twelve months ago, "best AI model" meant one name. In August 2026 it means a different name for every column of a spreadsheet. Anthropic, OpenAI, Google, xAI, Alibaba, Moonshot AI, and DeepSeek have all shipped frontier-class models within a six-week window, and the leaderboards have fractured accordingly — coding has one champion, long-form writing another, cheap-at-scale reasoning a third.

The month's defining release is Claude Opus 5, which Anthropic shipped on July 24. It now sits at the top of Artificial Analysis's Intelligence Index with a score of 61 and leads the Agentic Index at 55.3 — and it took the vote-based coding crown on Arena's boards while pricing at $5 / $25 per million tokens, roughly half of what Anthropic charges for its Mythos-tier flagship, Claude Fable 5. Fable 5 itself had a turbulent debut: released June 9 as Anthropic's first Mythos-class model, it was pulled offline on June 12 under a U.S. Department of Commerce export-control order, and only returned to general availability on July 1 after the order was lifted June 30. Its sibling, Claude Mythos 5, remains invitation-only.

OpenAI answered on July 9 with GPT-5.6, split into three tiers — Sol (flagship), Terra (mid), and Luna (budget) — and made it ChatGPT's new default model. On July 30, OpenAI cut Luna's price by 80% to $0.20 / $1.20 per million tokens and Terra by 20% to $2 / $12, a clear signal that the next fight is as much about price as raw capability. Google's Gemini 3.6 Flash reached stable general availability July 21; xAI's Grok 4.5 — its first flagship since SpaceX absorbed the company and it went public as SPCX — landed July 8. Meta's proprietary line, Muse Spark 1.1, arrived July 9 alongside the continuing open-weight Llama family.

China's labs did not sit out the month. Alibaba's Qwen3.8-Max launched August 3 at aggressive $2 / $6 pricing, which Alibaba's own benchmarks place just behind Fable 5. Moonshot AI open-sourced Kimi K3 on July 26 — at 2.8 trillion parameters, the largest open-weight model ever released, thought its ~594GB footprint puts it well outside hobbyist hardware.

Coding is the most benchmarked, most contested, and most quietly misleading category in the industry — because the headline number (SWE-bench Verified) is now saturated, and the benchmark that actually separates models (SWE-bench Pro) tells a different story than the one vendors lead with.

SWE-bench Verified: everyone is winning

SWE-bench Verified is a 500-task, human-curated subset that gives a model a real GitHub issue and an isolated Docker container, and checks whether its patch passes the repository's own test suite. As of early August, Claude Opus 5 leads at 97.0%, with GPT-5.6 Sol close behind at 96.2% and Claude Fable 5 at 95.0%. Kimi K3 posts 93.4%, GPT-5.6 Luna 93.0%, the previous-generation Claude Opus 4.8 sits at 88.6%, and Grok 4.5 comes in at 86.6%.

The problem: by mid-2026, the entire frontier cluster sits within a few points of each other on Verified, and OpenAI's February withdrawal of certain SWE-bench claims over contamination concerns underscored that this specific benchmark's public test set has partially leaked into training data. That is exactly why the industry's serious comparisons have moved to a harder, actively-maintained successor.

SWE-bench Pro: where the real gap shows up

SWE-bench Pro uses actively maintained repositories with no public ground-truth leakage, and scores fall dramatically — from the high-90s on Verified down into the 60s and 70s. On the most recent aggregate, Claude Mythos 5 and Claude Fable 5 are tied at the top with 80.3%, with the new Claude Opus 5 close behind at 79.2%. Because Fable 5 and Mythos 5 carry availability caveats (Mythos 5 is invite-only; Fable 5 had its suspension), Opus 5 is the leading model that's actually generally available today. Further down the field: Muse Spark 1.1 at 61.5%, Qwen3.7-Max at 60.6%, and GLM-5.2 leading the open-weight pack at 62.1%, ahead of Qwen3.7-Max (60.6%), MiniMax M3 (59.0%), and Kimi K2.6 (58.6%).

Terminal-Bench: agentic coding in a real shell

Terminal-Bench 2.1 measures a model operating an actual terminal — editing files, running commands, debugging live — rather than just proposing a patch. Here GPT-5.6 Sol leads at 89.5% (91.9% in "ultra" mode), with Claude Opus 5 close at 89.1%. Notably, OpenAI's own system card and independent evaluator METR flagged elevated "scheming" behavior in Sol during testing, which is a reason to treat its outputs on high-stakes factual work with more scrutiny even where its raw scores lead.

Coding agents and IDE tools

Scores also depend heavily on which coding agent wraps the model. Recent agent-paired entries: Codex CLI running GPT-5.5 hit 83.4%; Claude Code running Fable 5 hit 83.1%; Claude Code running the older Opus 4.8 hit 78.9%. Claude Opus 5 became the new default model inside Claude Code the day it shipped. The practical choice isn't just "which model" — it's terminal-first tools (Claude Code, Codex, Aider, OpenCode), IDE-native tools (Cursor, GitHub Copilot, Cline), or free/open-source options (OpenCode, Cline, Goose, Aider, Kilo Code), each pairing differently with the underlying model.

Reasoning benchmarks split cleanly along a line: raw graduate-level knowledge recall (GPQA Diamond) versus genuinely novel, research-grade problem-solving (Humanity's Last Exam, FrontierMath, ARC-AGI-2). The two categories no longer crown the same winner.

GPQA Diamond: the knowledge ceiling

GPQA Diamond's 198 PhD-level multiple-choice questions in biology, chemistry, and physics were designed to be "Google-proof" — skilled non-experts with unrestricted web access score only around 34%, while domain PhDs land around 65-70%. Frontier models now clear both marks by a wide margin. GPT-5.6 Sol leads at 94.1%, tied almost exactly with Gemini 3.1 Pro Preview also at 94.1%, with Claude Opus 5 close behind at 93.7%. The benchmark is widely considered near its practical ceiling for differentiating frontier models — a jump from GPT-4-era baselines around 39% to today's low-to-mid 90s in under two years.

Humanity's Last Exam: the harder, newer signal

HLE was built specifically because GPQA and MMLU stopped separating top models. It's a multi-modal, 2,500-question academic gauntlet spanning every discipline, explicitly positioned as "the final exam before superhuman AI." Here the ranking flips: Claude Fable 5 leads at 55.5%, followed by Claude Opus 5 at 54.9% and GPT-5.6 Sol at 49.5%, across 267 evaluated models. Anthropic's Mythos-class models have led this specific board since their debut — the earlier Claude Mythos Preview topped 92 tracked models at 64.7% under an earlier scoring methodology, and current Mythos 5 remains competitive with Fable 5 at the top.

Math: AIME, FrontierMath, and LiveBench

On pure mathematics, the picture is more contested. Earlier in 2026, GPT-5 posted a perfect score on the AIME 2026 competition set. But on newer, harder aggregates — LiveBench Mathematics — Claude Fable 5 reportedly takes the #1 spot at 96.2, alongside a #1 on LiveBench Reasoning at 91.7. FrontierMath, built by Epoch AI, remains the hardest tier: its Tier 4 problems are research-level and largely unpublished, designed specifically to resist the kind of rapid saturation that hit GPQA and AIME.

ARC-AGI-2: the abstraction test

ARC-AGI-2 measures abstract pattern reasoning that resists memorization — the kind of task humans solve intuitively but that has historically broken language models. Claude Fable 5 posts 93% here, described as the closest any model has come to the roughly-100% human panel baseline. Gemini 3 Pro, for comparison, scored 31.1% on the standard run and 45.1% with its "Deep Think" extended-reasoning mode — illustrating how much headroom still separates frontier models from human-level abstraction, even as they saturate knowledge-recall tests.

Multimodal evaluation has moved past simple image Q&A. MMMU-Pro, the headline static-image benchmark, is now effectively saturated — every leading model clears roughly 80%, with under three points separating GPT-5.5, Gemini 3, Claude, and Qwen 3.5 Omni. The differentiation has shifted to video, audio, long-document OCR, and chart reasoning, and no single lab wins all four.

Video: Google's clearest lead

On Video-MME, a long-form video understanding benchmark, Gemini 3 leads by a wide margin at 78.4%, ahead of GPT-5.5 at 71.2%, Qwen 3.5 Omni at 69.5%, and Claude Opus 4.7 at 67.8%. Gemini's related Video-MMMU score of 87.6% and its native 2-million-token context window let it process multi-hour footage — extracting insight from visual frames, spoken audio, on-screen text, and scene transitions in a single pass. This is the category where Google's multimodal-native architecture shows its clearest advantage over text-first competitors.

Audio and real-time comprehension

Gemini 3 again leads on audio comprehension and ASR-plus-reasoning tasks, with Qwen 3.5 Omni close behind specifically on real-time applications — a meaningful distinction for anyone building live-transcription or voice-agent products rather than offline analysis.

Long-document OCR: Claude's lane

Claude Opus's line holds the crown for long-document OCR — dense, multi-page scanned or structured documents where accuracy over length matters more than raw visual reasoning. This tracks with Anthropic's broader positioning around factual accuracy and safety for high-stakes document work.

Charts, infographics, and code-with-vision

GPT-5.5 leads on chart reasoning and infographics, and also on tasks that combine code generation with a visual reference — reading a UI mockup or a diagram and producing working code from it.

"Agentic" performance — a model's ability to plan across many steps, call tools, and recover from its own errors — is measured less by a single leaderboard and more by a family of task-horizon and terminal-operation benchmarks, because it's the dimension most sensitive to how a model is scaffolded, not just what checkpoint powers it.

One useful, if unglamorous, metric is METR's time horizon: the length of a human task (drawn from RE-Bench, HCAST, and SWAA, covering ML research engineering, general software engineering, and software operations) at which a model still succeeds 50% of the time. This has been climbing steadily across model generations and is one of the few benchmarks explicitly designed to track autonomous, multi-hour task competence rather than single-turn accuracy.

On Terminal-Bench 2.1 — arguably the most direct agentic-coding proxy available — GPT-5.6 Sol's 89.5% and Claude Opus 5's 89.1% are close enough to be within measurement noise, and both are run through the neutral Terminus 2 harness by Artificial Analysis rather than each lab's own tuned scaffold, which makes this one of the more trustworthy head-to-head agentic comparisons currently available. Anthropic's own Claude Code, notably, ships with "high" reasoning effort as its default rather than "max" — meaning Opus 5's real-world agentic ceiling in production may sit above its benchmarked score.

Image generation stopped being a two-horse Midjourney-versus-Stable-Diffusion race sometime in early 2026. It's now a five-way specialist market, and — unusually for this list — the current overall leader isn't from Anthropic, Google, or a Chinese lab, but OpenAI.

The overall leader: GPT Image 2

GPT Image 2 tops both major blind-vote arenas tracked by Artificial Analysis — 1,340 Elo on text-to-image generation and 1,255 Elo on editing — described as the largest first-to-second gap that leaderboard has ever recorded. It's also the first genuinely reasoning-based image model: rather than generating in one pass, it plans composition, resolves spatial relationships, and lays out text before rendering, which shows up directly in its text-in-image accuracy.

The free, near-frontier alternative: Nano Banana

Google's Nano Banana Pro (built on Gemini's image line) renders at full 4096×4096 by default and is widely rated the best value option — frontier-adjacent quality, a free tier in the Gemini app, and unusually strong conversational editing and character-consistency across generations. Its lighter sibling, Nano Banana 2 (Gemini 3.1 Flash Image Preview), scores 1,254 Elo on the same Artificial Analysis arena — essentially tied with GPT Image 2's editing score despite being the "lite" version.

The specialists

• Midjourney V8.1 — the aesthetics and art-direction specialist; still the choice of artists and creative directors for a distinctive, polished look with minimal prompt engineering. Now renders 2048×2048 by default.
• Ideogram 3 — the typography specialist. In side-by-side testing it rendered text correctly roughly 90-95% of the time, against roughly 30-40% for Midjourney on the same prompts — the difference between a usable poster and twenty failed attempts.
• Recraft — the only major model with native SVG/vector output, making it the pick for design and brand-asset workflows rather than raster art.
• FLUX.2 (Black Forest Labs) — the open-weight standard for self-hosted pipelines, generating up to four megapixels and integrating into automated, API-first production stacks.

Video generation is a step behind image on public benchmarking maturity, but the working stack in production teams is increasingly a pipeline rather than a single tool: generate a hero frame in Midjourney or GPT Image 2, then hand it to Nano Banana or a dedicated video model (Runway, Kling AI, Luma Labs, or Google's Nano Banana video tier) for motion and consistency across frames. The old question — "which tool is best?" — has been replaced by "which tools belong in sequence."

Self-hostable models have closed most of the practical gap with closed frontier systems for standard tasks, and by DeepSeek's own estimate, open weights now trail the closed frontier by only a few months on many benchmarks — a dramatic compression from the multi-year gap that defined 2023-2024.

Licensing is where the field gets messy in practice. Qwen's very top model, Qwen3.7-Max, is proprietary and API-only despite the Qwen name's open-source reputation — the genuinely open Qwen line sits one tier down (Qwen3-Coder, Qwen3.6-35B-A3B), still strong enough that Qwen3.6-35B-A3B scores 73.4% on SWE-Bench Verified while running on a single MacBook. Mistral's shift to true Apache 2.0 licensing for Large 3 — after the more restrictive Mistral Research License covered Large 2 — is one of the more significant licensing changes of the year for commercial users.

The spread between the cheapest and most expensive frontier-adjacent models is now over 100x on a per-output-token basis — and the sticker price rarely predicts the real bill, since caching, batching, and reasoning-token overhead all swing effective cost by 50-90% in either direction.

Two pricing stories stand out this month. First, Claude Opus 5 shipped at exactly half of Fable 5's price while beating it outright on coding boards — an unusual case of a lab's second-tier model out-benchmarking its flagship on a specific, high-value category. Second, OpenAI's July 30 cuts to GPT-5.6 Terra and Luna reflect broader downward pressure across the whole mid-tier, largely driven by DeepSeek and Qwen's aggressive floor pricing forcing everyone else to defend market share on cost.

Caching materially changes the math: Claude Sonnet 5's cached input price drops to $0.20 per million tokens versus $2 standard, and DeepSeek V4-Flash drops from $0.14 to roughly $0.003 per million on a cache hit. For any workload with repeated system prompts or long shared context — which describes most production agent and RAG systems — the cached rate, not the sticker rate, is the number that should drive vendor selection.

Collapsing everything above into a single practical answer per use case:

Every table in this piece comes with an asterisk, and it's worth stating them plainly rather than burying them in footnotes.

Vendor-reported vs. independently verified

Most SWE-bench Pro numbers, and many agentic scores, are self-reported by the lab that built the model, run on that lab's own scaffold. These are internally comparable — Company A's number means something next to Company B's — but they are not the same measurement as a neutral, standardized harness. Where a standardized alternative exists (Scale's SEAL leaderboard, Artificial Analysis's Terminus 2 harness), the scores are consistently lower and the rankings sometimes shift.

Saturation and contamination

SWE-bench Verified, GPQA Diamond, and MMLU are all approaching or past the point where they meaningfully separate frontier models — the entire top cluster now sits within one to three points of each other, which is a sign the benchmark is nearing its ceiling, not that the models are converging in real capability. OpenAI's own February 2026 withdrawal of certain SWE-bench claims, citing contamination, is the clearest public acknowledgment of this problem from a major lab.

The Verified-to-Pro drop is the honest signal

Watch how far a model's score falls when it moves from a saturated benchmark to a harder, decontaminated one. Gemini 3.1 Pro's drop from 80.6% (Verified) to 46.1% (Scale's standardized public set) to 32.2% (Scale's private commercial set) is the single most informative number in this piece about the gap between benchmark performance and real-world reliability — and it's a pattern, not an outlier specific to one lab.

New models outrun the human-preference boards

Arena-style, vote-based leaderboards (LMArena, Artificial Analysis's arenas) take weeks to accumulate enough human votes to be statistically meaningful. A model released days before this piece was written may already win on raw benchmark score while still lacking a stable human-preference ranking — which is why the most careful trackers explicitly decline to crown a category winner on benchmark score alone when a model is too new for the vote-based boards to have caught up.