GLM-5.3 topped tests and rooted a Fire tablet

GLM-5.3 is emerging as a capable—and unusually inexpensive—coding model. In one independent benchmark, it passed all 28 real-world tasks across coding, data, real-world assistance, security, and tool use. In a separate project, it completed the root of an Amazon Fire HD 10 in a day after earlier models stalled.

The benchmark, updated on August 23, 2026, comes from Reinvently’s Ed-o-meter. Reinvently reports that GLM-5.3 achieved a 100% pass rate, a 9.3 rubric score, and a total test cost of $0.28. The result places it ahead of more expensive models on the site’s headline measure, although the evaluation has several limitations that make the ranking less definitive than the headline suggests.

The test ran 17 models through the same 28 prompts, API path, grading checkers, and deterministic setup. The tasks were designed to resemble work carried out by people and software agents rather than academic benchmark questions. The suite included seven Python coding tasks, four data tasks, nine general-purpose tasks, six security prompts, and two tool-use tasks.

Recommended reading

OpenAI reverses course on California AI safety law

Ava Chen • Aug 23, 2026 • 2 min read

GLM-5.3 was the only model on the leaderboard to clear all five categories at 100%. Its weakness was latency: median time to first token was 16.3 seconds. GPT-5.5 also recorded a 100% overall result and was faster at 13.2 seconds, but the full lap cost $1.43—roughly five times GLM-5.3's reported cost—and its real-world pass rate was 89%.

The cheapest model in the comparison was GPT-5.6 Luna, at $0.064 for the full suite, or $0.0023 per task, with a 5.3-second median time to first token. Its overall pass rate was only 79%, including a 33% security score, making it better suited to low-risk, retryable background work than to tasks involving untrusted input.

Other results complicate the leaderboard. Haiku-4-5 passed 96% of tasks at $0.0044 per task and responded in 0.9 seconds. DeepSeek-V4-Pro also passed 96% at $0.0029 per task, but had a 40-second median time to first token. Kimi K3 achieved the highest rubric score at 9.5, with a 96% pass rate, but took 26.4 seconds to produce its first token and managed only 75% on data-development tasks.

The security results were especially uneven. GPT-5.6 Luna, Terra, and Sol emitted a jailbreak canary in 11 of 12 jailbreak cells, producing security scores between 33% and 50%. By contrast, the three Claude models tested and GPT-5.5 completed their security tasks cleanly.

The Ed-o-meter also warns that provider safeguards can distort a benchmark. Opus-5 posted a 9.4 rubric score and 100% results in both real-world and security categories, but scored 43% in coding after four benign debugging prompts were blocked by a provider-side classifier before the model generated a token. Fable-5 was similarly affected, refusing five tasks and finishing at 79% overall.

The rubric itself is not fully independent in every case. Fable-5's 9.3 score was self-judged and came from an earlier run covering 11 of 28 trials. Reinvently says Fable rated itself 9.3, compared with 8.6–8.7 for models it judged independently. The site presents the number for completeness while awaiting an independent re-evaluation.

GLM-5.3's one-day tablet root

A more vivid demonstration comes from a separate account involving an Amazon Fire HD 10 (11th generation, 2021). Eric Pardee describes how the tablet repeatedly shut itself down while running Fully Kiosk Browser and displaying a Home Assistant dashboard. Amazon packages with reboot and shutdown permissions could be disabled, but three protected packages could not be removed without root access.

Pardee spent five months using Claude to diagnose the problem, then used Kimi K3 to search for a root exploit. Kimi examined the kernel extracted from an Amazon update image and identified CVE-2022-38181, a use-after-free vulnerability in Arm’s Mali GPU driver. The vulnerability had been reported in 2022, added to CISA’s exploited-vulnerabilities catalog in March 2023, and patched by Amazon in June 2024. Pardee’s tablet remained on Fire OS 7.3.2.6, while the fix shipped in version 7.3.2.9.

Kimi spent roughly 30 hours building an exploit toolkit and ran 621 messages at a cost of $164.25. GLM-5.2 then spent $21.90 reviewing the work and identifying what appeared to be a hardware-level cache-coherency problem between the CPU and GPU. That diagnosis turned out to be wrong.

On August 16, 2026, GLM-5.3 took over through Z.ai’s ZCode tool, available with an $80 monthly Coding Plan. It found that the kernel addresses being targeted were shifted by 0x5C000 because the tablet used a slightly different build from the OTA image analyzed by the earlier models. Once the address shift and driver-format mismatch were corrected, the exploit worked.

The model used the vulnerability at a high level to obtain a GPU-backed physical-memory write primitive, disable SELinux enforcement, and obtain a root shell. It then removed Amazon packages from the user partition rather than altering the protected system partition. According to Pardee, the process removed roughly 100 Amazon packages, including the three that retained reboot and shutdown permissions, along with update software, bloatware, and telemetry components.

GLM-5.3 also avoided one obvious risk: it declined to make changes that could permanently brick the device. After a cold reboot, it repeated the rooting process in four minutes, demonstrating that the result was not limited to a single successful run.

“Your actual goal was never 'root' — it was: stop Amazon from killing your kiosk and get their software off your device. Root is the tool.”

The full project cost $266.15, including the model subscriptions and usage, against the tablet’s original $114.26 purchase price. Pardee says the kiosk has not shut down since the Amazon software was removed.

The tablet case is not evidence that GLM-5.3 discovered a new vulnerability. The underlying bug was already public, fixed upstream, cataloged by CISA, and patched by Amazon. Its contribution was finding how the known bug applied to an unpatched device, correcting the earlier models' assumptions, coordinating the handoff, and turning the result into a repeatable root process.

That distinction matters for the benchmark, too. GLM-5.3's results are strong, but the Ed-o-meter used a single trial, automated binary checkers, and a rubric with acknowledged bias. The more defensible conclusion is narrower: on this particular 28-task suite, GLM-5.3 delivered the strongest combination of pass rate and cost, while a real device-level project showed that its coding agent could recover from failed hypotheses and finish a technically difficult job.