Grok 4.6, GPT-5.6-Cyber, DeepSeek V4 Pro: AI News
Grok 4.6 benchmarks, GPT-5.6-Cyber use cases, DeepSeek V4 Pro, Omarchy Quattro RC, OpenAI departures, and T3 Code PR tabs for August 2026.
This AI developer news update covers six changes from August 10 through August 13, 2026. Grok 4.6 and DeepSeek V4 Pro are now available. OpenAI released GPT-5.6-Cyber through a controlled security program. T3 Code added pull request tabs to its nightly build. Omarchy Quattro reached its first release candidate today, but the final release is not available yet.
This post was verified on August 13, 2026. Release candidates, nightly builds, and model roadmaps can change after this date.
The news at a glance
| Date | Release or change | Main impact |
|---|---|---|
| Aug 10 | GPT-5.6-Cyber | A specialized model for approved, advanced cyber defense research |
| Aug 10 | T3 Code PR tabs | Pull request work moves into the agent control plane |
| Aug 12 | Grok 4.6 | Better long-running agent work at the same API price as Grok 4.5 |
| Aug 12 | DeepSeek V4 Pro GA | The larger V4 model moves from preview to API and chat availability |
| Aug 13 | Omarchy Quattro RC | The Arch-based developer distribution starts final public testing |
| Recent months | OpenAI safety leadership changes | Three senior ethics, safety, and alignment leaders have left |
Grok 4.6 focuses on long-running agents
SpaceXAI released Grok 4.6 on August 12. The release builds on Grok 4.5. Its main target is long-running agent work across coding, research, visual tasks, and interactive applications.
Grok 4.6 costs $2 per million input tokens and $6 per million output tokens at the standard API rate. This is the same base price as Grok 4.5. It launched in Grok Build, Cursor, the SpaceXAI API, and partner platforms. SpaceXAI also announced twice the normal usage in Cursor and Grok Build for the first week.
The first benchmark results
The first results show progress in agent and knowledge-work tasks.
| Benchmark | Grok 4.6 result | Useful comparison |
|---|---|---|
| GDPVal-AA v2 | 1753 Elo | Fable 5 Max: 1741; GPT-5.6 Sol Max: 1728; Grok 4.5 High: 1526 |
| CursorBench 3.2, Extra High | 70.8% | Fable 5 Max: 70.5%; Opus 5 Max: 70.0%; GPT-5.6 Sol Max: 67.2% |
| DeepSWE 1.1 | 65.9% | Grok 4.5: 54.0% |
The GDPVal-AA leaderboard shared after launch puts Grok 4.6 first in that evaluation. CursorBench also reports a cost of $2.81 per task and 46 steps for the Extra High run. These are useful signals, but they do not prove that one model is best for every repository. The agent harness, tools, prompts, and verification loop affect the result.
The practical claim is more important than a small leaderboard lead. SpaceXAI says Grok 4.6 tests more of its own work and stays with complex projects for longer. Test that claim on a completed task from your own repository. Measure successful completion, total cost, elapsed time, and required human corrections.
SuperGrok received a usage reset
A separate SuperGrok notice on August 12 said that all SuperGrok users received a free usage reset. The one-time reset cleared the current weekly usage and started a new allowance. It did not stack with an unused reset.
This was a reset granted yesterday. It was not an offer that remained open until August 12.
Grok 4.7 is already on the roadmap
Elon Musk said that initial Grok 4.7 training is complete. He expects the model in three to four weeks after supplemental training with SpaceX company data. Treat this as a roadmap estimate, not a fixed release date.
My take: Grok 4.6 is most interesting for long coding and research runs. Its API price and early cost-per-task numbers make it a serious model to evaluate. Do not select it from one benchmark table. Run it in the same agent and on the same tasks as your current model.
GPT-5.6-Cyber is for approved advanced security work
OpenAI released GPT-5.6-Cyber on August 10. It is built on GPT-5.6 Sol. It is not a general ChatGPT model and it is not the correct default for normal software work.
The model is available only through Daybreak Red. This controlled program is for approved individuals and organizations that perform authorized vulnerability research, exploit validation, and advanced security testing. OpenAI applies identity checks, account controls, monitoring, and legal restrictions.
OpenAI now separates cyber access into two paths:
| Access path | Model access | Use it for |
|---|---|---|
| Daybreak Blue | Frontier general models, including GPT-5.6 Sol, with defensive safeguards | Secure code review, vulnerability discovery, malware analysis, incident response, and patch validation |
| Daybreak Red | GPT-5.6-Cyber and purpose-built cyber models | Authorized zero-day research, exploit validation, exploit-chain analysis, and advanced red-team work |
Most security teams should use Daybreak Blue. Use Daybreak Red only when the work requires advanced exploit capability and your organization has clear authorization, isolation, review, and disclosure processes.
What changed in the cyber model
GPT-5.6-Cyber was trained for zero-day discovery and exploit-chain development. OpenAI reduced some refusals for high-risk dual-use tasks inside the controlled program. This makes the model more capable, but it also requires a narrower access boundary.
On OpenAI's internal Advanced Cybersecurity Completion Rate, GPT-5.6-Cyber scored 95.0%. GPT-5.5-Cyber scored 57.3%. Standard GPT-5.6 Sol scored 1.5%, and Sol with Daybreak Blue scored 2.0%. This evaluation measures the ability to complete advanced cyber tasks. It does not measure general secure-code-review quality.
OpenAI used the model to find two vulnerabilities in the V8 JavaScript engine. The model chained them into a heap sandbox escape. The issues were disclosed and fixed. One received the identifier CVE-2026-15903.
Use GPT-5.6-Cyber in a sandbox. Limit network, credential, and repository access. Require human review before a result reaches a production system. Use the standard GPT-5.6 Sol model with Daybreak Blue when exploit generation is not required.
My take: GPT-5.6-Cyber is not a better coding model for ordinary teams. It is a specialist tool for approved security experts. The access controls are part of the product, not a separate policy layer.
DeepSeek V4 Pro reaches general availability
DeepSeek V4 Pro entered general availability on August 12. The rollout covers the API and DeepSeek Chat. The V4 family first appeared as a preview in April.
DeepSeek V4 has two main variants:
| Model | Total parameters | Active parameters | Context window | Best fit |
|---|---|---|---|---|
| DeepSeek V4 Flash | 284B | 13B | 1 million tokens | Lower-cost and faster agent work |
| DeepSeek V4 Pro | 1.6T | 49B | 1 million tokens | Hard coding, reasoning, and long-context work |
DeepSeek says the family was pre-trained on more than 32 trillion tokens. The V4 Pro model card provides downloadable weights under the MIT license. This gives teams a route to controlled deployment, but the 1.6-trillion-parameter model still needs substantial infrastructure.
Use V4 Pro when quality on a hard task is more important than latency. Use V4 Flash when throughput and cost are more important. For either model, test tool use, long-context recall, and patch quality on your own agent harness before a production rollout.
My take: General availability makes V4 Pro easier to evaluate. The open weights also provide an alternative to closed frontier APIs. Most teams should start with the hosted API because self-hosting the full model is an infrastructure project.
Omarchy Quattro has an RC, not a final release
DHH published the first Omarchy Quattro release candidate on August 13. The final release is planned for Friday, August 14, if the release candidate passes testing.
Omarchy is an opinionated Arch Linux distribution for developers. It combines Hyprland, terminal tools, applications, themes, and system setup into one maintained environment. Quattro is its largest internal change so far.
The Quattro release change set includes these main changes:
- A new desktop shell built with Quickshell replaces several separate processes for the bar, launcher, notifications, lock screen, idle control, background, and policy prompts.
- Omarchy moves from a Git-based update flow to Arch packages managed by Pacman and its own package repository.
- The installer can set up dual boot when it finds free disk space.
- The ISO is more than 1 GB smaller, and installation is about 30% faster.
- Hyprland configuration moves to Lua for Hyprland 0.56.
- Theme colors expand from 8 to 24. Omarchy can generate themes for btop, Neovim, and VS Code.
- NetworkManager replaces the previous network setup.
- The launcher gets nested search, clipboard history, emoji search, controls, and other shell functions.
- Users can select a default coding agent. The system also remembers window positions.
The upgrade is expected to appear under Update > Omarchy after the final release. Existing custom Quickshell changes can require manual work because the shell architecture changed. Test the release candidate on a non-critical machine or wait for the final package.
My take: The package-based update system is the most important long-term change. The single Quickshell process can also reduce integration drift. The new shell is a large change, so the release-candidate label matters.
Three OpenAI ethics and safety leaders have left
A widely shared August 11 post grouped three recent OpenAI departures:
- Chloé Bakalar, Head of Ethics, left less than one year after joining.
- Johannes Heidecke, Head of Safety Systems, left after OpenAI reorganized that function.
- Josh Achiam, who had led Mission Alignment, reportedly left after the team was disbanded.
These departures did not occur as one coordinated announcement. The timeline covers several months. OpenAI disbanded the Mission Alignment team earlier in 2026. The company later reorganized Safety Systems, and Heidecke's departure was reported in July.
The leadership changes deserve attention because they affect three related functions. They do not, by themselves, prove that OpenAI ended its safety work. OpenAI continues to publish safeguards and to place controlled access around models such as GPT-5.6-Cyber. There is also no public evidence that the departures caused the cyber release.
My take: Watch who owns these functions next. Team structure, authority, and published evaluation data are better signals than a viral list without dates.
T3 Code adds PR tabs and two weeks of agent features
T3 Code PR tabs went live in the nightly build on August 10. T3 Code is an open-source control plane for coding agents. The new pull request surface keeps review work beside agent threads, repository state, and diffs.
The work from July 30 through August 13 includes more than a new tab:
- Multi-provider pull request pages and in-app reviews.
- Pull request filters, update-branch actions, reactions, inline editing, and improved diffs.
- Thread search across conversation content, not only thread titles.
- Native status for subagents and workflows, including a running-subagent count.
- A per-project choice between a new worktree and the current checkout.
- Mobile usage views and an hourly usage view for the last 24 hours.
- Open VSX theme search and a larger theme system.
These features are moving through a nightly channel. Expect small changes before they reach the stable release. The PR tab is useful now for teams that review agent work from the same tool, but GitHub remains the source of truth for required checks and final merge state.
My take: T3 Code is becoming a full agent workbench. The PR surface, conversation search, and subagent status reduce context switching. The best feature is not one more model selector. It is one place to see what the agents changed and what still needs human review.
What to test today
- Run Grok 4.6 on one known coding task and compare completion cost, steps, and corrections.
- Use GPT-5.6 Sol with Daybreak Blue for normal defensive security work. Apply for Daybreak Red only for authorized advanced research.
- Test DeepSeek V4 Pro through the hosted API before planning self-hosted infrastructure.
- Use Omarchy Quattro only as a release candidate today. Wait for the final release on a primary workstation.
- Try T3 Code PR tabs in nightly if your team accepts nightly-build risk.
The common theme is execution. Model vendors now compete on long-running work, controlled specialist access, and cost per completed task. Developer tools now compete on the control plane around that work. Benchmarks help with the first filter. A repeatable test on your own repository must make the final choice.