Grok 4.7 is now available in Kiro
Model choice is a core part of how we build Kiro. With Grok 4.7, Kiro now supports models from Anthropic, OpenAI, SpaceXSI, and leading open weight providers, all in one place. A diverse lineup lets you pick for capability, speed, and cost on each task, and keeps your workflow steady as the frontier keeps moving.
Grok 4.7 from SpaceXSI is available today in Kiro IDE, CLI, Crew, and Web. SpaceXSI calls it its most capable model for coding and knowledge work. It’s built for stamina, staying on hard, multi-hour tasks and checking its own work before it moves on.
Video transcript
Step-by-step transcript: A purple title card reads “Grok 4.7 now available in Kiro” next to the Kiro ghost mascot. A table titled “Grok 4.7 vs Grok 4.6” from Artificial Analysis follows, under the heading “A bigger step up for coding agents”: Intelligence Index 44 to 46, Coding Agent Index 47 to 56, AA-Briefcase long-horizon work 1,546 to 1,657 Elo, GDPval-AA professional work 1,605 to 1,695 Elo, AA-Omniscience Index 30 to 32, AA-Omniscience hallucination rate 34% to 29%, and output tokens per Intelligence Index task about 38k to about 81k, citing the AWS Machine Learning Blog. Kiro IDE, labeled “Long coding tasks, spec by spec,” shows a spec’s implementation plan in tasks.md with Grok 4.7 selected in the chat panel as it works through the first task. A DeepSWE v1.1 bar chart, headed “Frontier-level fixes at a fraction of the price,” shows GPT-5.6 Sol at 72.7%, Grok 4.7 (high) at 71.0% labeled “+5.8 pts over Grok 4.6,” Fable 5.1 at 70.0%, and Grok 4.6 at 65.2%. Kiro Web, labeled “Hand off fixes, get PRs back,” shows a session that pushed security changes to several repositories and lists a pull request for each. A GDPval professional knowledge work chart, headed “Self-validation over long sessions,” shows Fable 5.1 (max) at 1,735 Elo, Grok 4.7 (xhigh) at 1,695 labeled “+90 Elo over Grok 4.6,” Grok 4.6 (high) at 1,605, and GPT-6 Astra (max) at 1,542. Kiro Crew, labeled “Multi-hour work after you log off,” shows a long-running session, opens the model picker where Grok 4.7 is marked “New,” and then shows Grok 4.7 selected alongside a list of changed files. Kiro CLI, labeled “Terminal work and automation,” shows the Kiro CLI V3 welcome screen, the model set to Grok 4.7, and a prompt to recover a broken PyTorch model. The video ends on a purple card reading “Ship quality software. Get started now at kiro.dev,” then the Kiro logo.
SpaceXSI trained Grok 4.7 on a new, larger base model, with a longer reinforcement learning run weighted toward tasks that take many hours and made it better at verifying its own output and at using its full context on long tasks. The gains show up most on long coding and terminal work. Per SpaceXSI’s published results, compared with Grok 4.6:
| Benchmark | Grok 4.6 | Grok 4.7 |
|---|---|---|
| Terminal-Bench 4.0 | 20.3% | 37.6% |
| CursorBench 4.0 | 40.4% | 46.3% |
| DeepSWE v1.1* | 65.2% | 71.0% |
Independent testing points the same way. Artificial Analysis, which runs its own evaluations, measured Grok 4.7’s Coding Agent Index at 56, up from 47 for Grok 4.6. It also measured the hallucination rate on its AA-Omniscience test falling from 34% to 29%.
Self-checking matters most on long runs, where one early mistake can carry through every later step. That fits spec-driven development in Kiro. When you hand Grok 4.7 a spec with a long task list, it’s built to confirm each step holds before it starts the next one.
- 01Read the next spec task
- 02Write the change
- 03Check its own work
- 04Move to the next task
repeat for every task in the spec
Reach for Grok 4.7 on long, context-heavy work where price matters. SpaceXSI’s own comparison against GPT-5.6 Sol and Fable 5.1 shows where it stands out:
| Use case | Why Grok 4.7 fits (SpaceXSI data) |
|---|---|
| Hardware and electrical engineering | Top score on EEBench at 64.0%, ahead of Fable 5.1 (56.4%) and GPT-5.6 Sol (39.4%) |
| Long coding tasks | 46.3% on CursorBench 4.0, ahead of GPT-5.6 Sol (41.7%), at $2 input / $6 output per million tokens |
| Software engineering fixes | 71.0% on DeepSWE v1.1, in the same range as GPT-5.6 Sol (72.7%) and Fable 5.1 (70.0%) at a fraction of their price |
Grok 4.7 has four reasoning effort levels: low, medium, high, and xhigh. The default is high. Pick the level based on intended purpose. Higher effort gives Grok 4.7 more room to check its work across a long task. It has a 500K token context window, so a full spec, its design doc, and a large slice of your codebase can sit in one session. It accepts text and image input. Use low for quick lookups and small edits. Save high and xhigh for multi-step plans, long agent runs, and bugs that span many files.
- lowLookups and small edits
- mediumEveryday coding
- high (default)Multi-step plans
- xhighLong agent runs and cross-file bugs
According to SpaceXSI, Grok 4.7 ships with an entirely new safeguard stack and is the strongest model SpaceXSI has tested on refusals and jailbreak resistance. On SpaceXSI’s HackerBench v0.3, it let only 3.3% of risky dual-use cyber prompts through, while rarely blocking legitimate security work.
Grok 4.7 is available today across Kiro CLI, Crew, Web, and IDE with experimental support for Kiro Pro, Pro+, Pro Max, Power, and Enterprise plan customers.
It’s available in us-east-1 (N. Virginia) and uses a 2.7x credit multiplier. For long context requests over 200K tokens, the multiplier doubles to 5.4x.