AI Model Tracker: July 22-29 Releases and Benchmarks

Published July 29, 2026 · Research window: July 22-29, 2026 (UTC) · Author context: Essa Mamdani, AI/full-stack developer

This week produced four distinct signals: vendor release claims, vendor-reported benchmarks, an independent math update, and an independent cyber assessment. Keep those evidence classes separate or the rankings become meaningless.

The week at a glance

Claude Opus 5: stronger, but still vendor-reported

Anthropic’s Opus 5 launch is the week’s clearest premium-model story. The official page says Opus 5 is the new default on Claude Max and the strongest model on Claude Pro, with vendor-reported wins on Frontier-Bench v0.1, CursorBench 3.2, AutomationBench, and OSWorld 2.0.

This screenshot proves the operational part: a model launch matters when it lands in a real workflow surface, not just a vendor demo.

Anthropic’s benchmark page is still vendor-reported, so treat Opus 5 as a strong candidate for difficult coding and review, not as a universal winner.

Grok 4.5: the launch that matters because of where it lands

xAI’s own docs and GitHub’s changelog make Grok 4.5 interesting for a different reason. xAI describes it as a frontier model for coding, agentic tasks, and knowledge work, with a 500,000-token context window, text and image support, and configurable reasoning effort. The docs list $2 per million input tokens and $6 per million output tokens. GitHub adds the deployment signal: Grok 4.5 is rolling out in Copilot for Pro, Pro+, Max, Business, and Enterprise users.

The distribution path matters as much as the model specs. A model that is only in a console is a release note; a model that appears in Copilot, VS Code, Copilot CLI, JetBrains, and Xcode becomes a workflow option.

For teams evaluating it, the useful question is whether the model stays useful when tool calls get messy and the agent has to recover from partial failure without losing the thread.

AIME26: the math leaderboard is bunched up

BenchLM’s July 28 AIME26 update is a reminder that frontier math is getting crowded. GLM-5.2 leads at 99.2%, Inkling follows at 97.1%, and Kimi K2.6 sits at 96.4%. BenchLM says the top three are separated by 2.8 points.

This is the only defensible numeric chart in the piece because all three bars use the same benchmark and the same reporting surface. It shows why math leaderboards are useful for screening but weak as a full product-selection rule.

If you are deciding whether a model is worth piloting, AIME26 is a signal, not a strategy.

Kimi K3: useful cyber signal, not a free pass

The UK AISI / CAISI assessment of Kimi K3 is the week’s strongest independent safety signal. The report says Kimi K3 scores 32% on ExploitBench versus 24% for GLM-5.2, reaches step 17 on The Last Ones range on average, completes it in 1 of 10 attempts, and gets 0 of 41 ExploitBench tasks to arbitrary code execution.

That last detail matters. A model can look strong on coding or math and still fail to cross the threshold that matters most in exploit development. The report also makes the methodology caveat explicit: these are preliminary results on a limited set of public and private benchmarks.

So the right reading is not “Kimi K3 is safe” or “Kimi K3 is dangerous.” Cyber capability is measurable, and it is bounded by benchmark design, configuration, and scenario.

What I would pick

For complex coding and review work, Opus 5 looks like the safest vendor-reported premium candidate this week.

For fast agentic workflows where deployment surface matters, Grok 4.5 is the one to watch because it is already landing in GitHub Copilot and xAI’s own API/docs.

For a reality check on reasoning and benchmark saturation, AIME26 is useful only if you remember that the top band is crowded.

For security-sensitive workflows, the Kimi K3 assessment is a reminder to run your own controlled evals before you trust any model near exploit-sensitive tasks.

If you are routing models across tools and surfaces, pair this tracker with AI Coding Agents: Skills, Plugins, MCP, ACP, Memory and Workflows, Structured Outputs for Reliable AI APIs, and OpenTelemetry GenAI observability for the implementation side.

FAQ

Are these scores directly comparable?

No. Claude Opus 5 and Grok 4.5 are vendor-reported release claims, BenchLM is an independent benchmark leaderboard, and the UK AISI / CAISI report is an independent cyber assessment.

Why include a vendor release with a benchmark tracker?

Because developer tooling changes fast when a model lands in a real product surface. A release in Copilot is not the same thing as a release in a blog post.

Does a 500K-token context window guarantee better long-context results?

No. It only means the model can accept long inputs. Retrieval quality, tool handling, latency, and truncation behavior still decide whether long-context work is usable.

Is AIME26 enough to pick a model?

No. It is good for screening top mathematical reasoning models, but it does not tell you how a model handles code review, tool calls, or production failure modes.

Should I treat Kimi K3’s cyber report as a safety verdict?

No. It is a narrow assessment of cyber capability, not a universal safety rating.

Methodology and scope

I searched official vendor, GitHub, benchmark, and government pages published in the July 22-29, 2026 window. Vendor release pages and model docs are treated as vendor-reported. BenchLM is independent. The UK AISI / CAISI assessment is independent. I did not merge those evidence types into a single rank.

I also avoided secondhand write-ups when an official source existed. For Grok 4.5, xAI’s docs plus GitHub’s changelog were the primary sources. For Opus 5, the official Anthropic release and GitHub Copilot availability note were the primary sources. For AIME26 and Kimi K3, the independent benchmark and government report were the right evidence classes.

Source notes

• Anthropic: Introducing Claude Opus 5
• GitHub: Claude Opus 5 is now available in GitHub Copilot
• GitHub: Grok 4.5 is now available in GitHub Copilot
• xAI docs: Grok 4.5
• xAI docs: Release notes
• BenchLM: AIME26 Leaderboard & Scores
• NIST: UK AISI / CAISI Preliminary Assessment of Kimi K3's Cyber Capabilities

Author context

I write about AI systems and full-stack engineering from the point where product behavior meets production reality. My bias is toward explicit contracts, narrow tool surfaces, and enough telemetry to explain failures without exporting private data everywhere.

If you want help turning this kind of tracker into a production-grade eval harness, I’m available through /hire.