GLM-5.2 just aced a task many accountants treat as routine drudgery. It prepared a quarterly VAT return for a small UK business. The net position missed the ground truth by seven pence. Raw token cost came to $2.73. Time taken clocked in at 68 minutes.
Adam Kurkiewicz ran the test at Vineyard Finance. He published the results on toot-books.pages.dev. The model processed 59 transactions. It used tools on a GCP instance via Fireworks AI. And it delivered output that stands up against a competent human bookkeeper.
But this isn’t an isolated win. GLM-5.2 from Z.ai has climbed leaderboards across reasoning, coding and agentic tasks since its June 2026 release. The model carries roughly 744 billion parameters in a mixture-of-experts setup. Only about 40 billion activate per token. That design keeps inference costs manageable. It ships with a one-million-token context window. MIT licensing lets developers run it freely.
Nearly perfect. That’s how Kurkiewicz described the VAT output. Out of 354 scored checks across six criteria per transaction, the model failed just 20. Those errors spread over 18 transactions. One stood out as serious. The model classified founding shares as a capital account instead of unpaid shares. Legal implications follow from that choice.
Fourteen failures involved confusion between zero-rated and exempt VAT treatments. The pattern appeared in January and February transactions but vanished in March. Three more stemmed from split transactions with Wise balances that led to double-counting. Still, the model nailed account classification in nearly every case. It attached the correct invoices without error. It handled ambiguous transfers and same-amount transactions with precision.
The benchmark itself drew from real 2026 Q1 books at Vineyard Finance. Claude Fable 5 helped extract the test data. The evaluation scored the final state inside accounting software. Six criteria applied to each transaction: type, category, VAT treatment, VAT amount, reverse-charge VAT and receipt attachment. Amounts tolerated variance of two pence.
Wall Street has watched Chinese labs close the gap for months. The Economist noted in June that America’s lead in artificial intelligence may sit at its narrowest in over a year. GLM-5.2 contributes to that pressure. It scores 51 on Artificial Analysis’s Intelligence Index. That places it ahead of previous open-weight leaders and competitive with closed models from OpenAI and Anthropic.
Developers have taken notice. On X, users report GLM-5.2 passing complex git workflows that involve dirty worktrees, commit history matching and rebasing without breakage. One engineer tested it on a 50,000-line codebase. It succeeded where others stumbled. Another compared it in OpenBench evals against Kimi 2.7 and DeepSeek variants. Pi harness paired with GLM-5.2 stood out in several agent setups.
Coding benchmarks tell a similar story. The model posts 62.1 percent on SWE-bench Pro, up from 58.4 for its predecessor. Terminal-Bench 2.1 delivers 81.0 percent. FrontierSWE reaches 74.4 percent. These figures put it within a few points of Claude Opus 4.8 on some measures while running at roughly one-sixth the cost, according to emergent.sh.
Semgrep researchers tested GLM-5.2 on cybersecurity tasks. It scored 39 percent F1 on IDOR detection using nothing but a bare prompt. That beat Claude Code at 32 percent. Cost per vulnerability found landed around 17 cents. The open-weight model outperformed a frontier coding agent in that setup. Semgrep’s analysis highlighted how the model’s long context helps trace authorization logic across files.
Yet benchmarks carry limits. Some observers on Hacker News point out that models from Chinese labs sometimes show wider gaps between public leaderboards and private evaluations designed to resist memorization. Contamination remains a concern across the industry. Real-world scaffolding still determines whether these scores translate to production value.
Pricing adds another layer. Providers on OpenRouter list GLM-5.2 from 95 cents to three dollars per million input tokens. Output ranges from three to more than ten dollars. Median sits around $1.40 input and $4.40 output per DeepInfra’s comparison. Cache hits in the VAT test reached 92 to 95 percent. That helped contain expense.
Z.ai recommends maximum thinking mode for coding work. The model offers two modes. Context reliability across one million tokens matters for repository-scale tasks or multi-step agents. Early users report the window holds up better than many competitors on long trajectories.
Financial applications could shift fastest. Bookkeeping has resisted full automation for decades. Regulatory nuance, edge cases and audit trails create friction. Kurkiewicz argues the VAT benchmark shows the problem is quickly becoming solved. Startups can now focus on scaffolding that wraps these models for UK SMEs. His team at toot-books.com builds exactly that.
Enterprises outside finance have started pilots too. Fireworks AI customers moved internal agents from Opus 4.8 to GLM-5.2. One partnership lead said nobody noticed the switch. Outputs stayed consistent. Token usage for similar workloads dropped sharply. Fireworks optimized the serving layer for throughput.
Competition intensifies. MiniMax M3 and DeepSeek V4 Pro trail GLM-5.2 on several coding lists but win on pure efficiency in some provider tests. Closed models from Anthropic and OpenAI still lead on select reasoning suites. The gap, however, has narrowed enough that cost and licensing tilt decisions toward open weights.
GLM-5.2 also shows strength on visual design leaderboards. It reportedly surpassed Claude on certain web design metrics. Multimodal training appears to help there. For VAT work the model operated in text and tool mode. It read transaction lists, ledger states and invoice images when attached.
Analysts expect more domain-specific benchmarks to emerge. The VAT test stands apart because it uses live business data instead of synthetic exams. Errors carry real consequences. A seven-pence miss on a return won’t trigger penalties. A ten-thousand-pound misclassification on share capital could.
So what comes next. Teams that integrate these models into accounting workflows may cut costs dramatically. Accuracy at human levels for fractions of a percent of prior expense changes the math. But oversight remains essential. The serious error on founding shares reminds that models still miss context humans absorb instinctively.
Industry insiders watch how regulators respond. UK HMRC rules around VAT have tightened. Automated filings already face scrutiny. If models like GLM-5.2 become standard, audit trails and explainability will matter more than raw accuracy.
The release also underscores China’s momentum. Z.ai spun from Tsinghua University. It operates under the Zhipu AI umbrella. Government support for domestic AI has produced a string of capable open models. Western firms now deploy them at scale. Usage on some platforms jumped 27 times in a single week after recent updates.
Developers on X continue to share novel tests. One ran GLM-5.2 against frontend coding arenas. It matched top performers. Another paired it with agent harnesses for multi-hour engineering projects. Results vary by prompt quality and scaffolding. The model rewards careful instruction.
Yet the VAT result may prove most telling. It targets a mundane but universal business process. Success there signals broader potential in compliance, reconciliation and reporting. Firms that once budgeted thousands per quarter for bookkeeping staff or outsourced services now face new options.
Kurkiewicz ended his post on an optimistic note for builders. Bookkeeping, he wrote, is becoming solved. The focus shifts to products that make these capabilities accessible to every SME. Others in financial technology will likely follow.
GLM-5.2 won’t replace every accountant tomorrow. It does demonstrate that frontier-level performance on practical financial tasks has arrived at commodity prices. For an industry that prizes both precision and efficiency, that combination demands attention.
WebProNews is an iEntry Publication