THE SHORT VERSION
● LongCat-2.5 — 1.6T total parameters, ~48B active, 1M context, 128K max output. $0.30 in / $1.20 out per million tokens.
● Momentic QA Bench — 1,396 labelled bugs. Mo matched 250.3 in 60 minutes at $2.88 each. Codex matched 94.3 at $0.56.
● GDPx — SemiAnalysis agent benchmark. Claude Opus 5.5 leads at 12.0%. Finance and legal score zero.
● Abu Dhabi — four OpenAI agreements, including a five-year Inception42 rollout. No values disclosed.
Meituan releases LongCat-2.5 with a 1M-token context
Meituan released LongCat-2.5 on 2 November 2026. It is a mixture-of-experts model with 1.6 trillion total parameters and roughly 48 billion active per token, natively multimodal, with a 1 million-token context window and a maximum output length of 128,000 tokens. Meituan positions it for long-horizon agentic work across terminals, browsers, GUIs, spreadsheets and design tools.
Pricing on Meituan’s own platform documentation is $0.30 per million uncached input tokens, $0.006 per million cached input tokens and $1.20 per million output tokens, listed as a limited-time discounted rate. The API speaks both the OpenAI and the Anthropic request formats, which means it drops into an existing Claude Code or OpenAI-compatible harness by changing a base URL rather than rewriting a client.
Weights are not open yet. Meituan says GitHub and Hugging Face releases are planned to follow. For now the model is reachable through Meituan’s chat service and its API platform. The combination that matters here is the 1M window and the 128K output ceiling at $1.20 per million output tokens — that is the configuration long agent runs actually bill against.
Momentic publishes QA Bench, a 1,396-bug benchmark
Momentic (YC W24) published QA Bench, which measures how many real bugs an agent finds when pointed at a running web application. The suite covers 1,396 human-labelled bugs across four open-source apps — Gitea, Metabase, Twenty and Saleor — spanning 20 product areas. Each configuration gets a fixed time budget rather than a fixed token budget.
Read this one with the conflict in view. Momentic built the benchmark, and Momentic’s own agent tops it. The figures below are the vendor’s, from exports taken 31 October 2026. We have not reproduced them.
| Configuration | Bugs matched (60 min) | Recall | Cost per bug |
|---|
| Mo (Momentic) | 250.3 | 17.9% | $2.88 |
| Codex, GPT-6.1 Sol, Playwright MCP | 94.3 | 6.8% | $0.56 |
| Codex, GPT-6.1 Terra, Playwright MCP | 72.7 | 5.4% | $0.52 |
| Claude Code, Opus 5.5, Playwright MCP | 39.6 | 2.8% | $2.79 |
| Devin | 27.0 | 1.9% | $3.89 |
| Claude Code, Sonnet 5.5, Playwright MCP | 8.0 | 0.6% | $1.40 |
The precision numbers cut the other way. Mo reported 84.7% precision at 60 minutes, second to Codex on GPT-6.1 Sol with Playwright MCP at 89.0%. So the cheapest configuration per bug is also the most accurate per report filed — it simply files far fewer. Mo’s recall was highest on Saleor at 31% and lowest on Metabase at 9%, which is a reminder that these numbers move with the application under test.
SemiAnalysis launches GDPx and nothing clears 12%
SemiAnalysis published GDPx, an agent benchmark built around tasks drawn from real economic work across 11 domains. Claude Opus 5.5 leads at 12.0%, Gemini 4 Carbon follows at 10.9% and GPT-6 Astra at 9.6%. SemiAnalysis says the benchmark refreshes every three to six months.
The domain breakdown is the interesting part: finance and legal tasks scored 0% for every model tested. A benchmark where the leader completes roughly one task in eight is measuring headroom rather than capability, which is the point — saturated benchmarks stop telling you anything about which model to pick for a given job.
Abu Dhabi signs four agreements with OpenAI
Abu Dhabi signed four separate agreements with OpenAI covering ADNOC, G42’s Inception42, Mubadala and the Abu Dhabi government itself, as part of a stated push toward AI-native government services. The Inception42 arrangement is a five-year rollout deal. No contract values were disclosed for any of the four.
Tutor Intelligence raises $75M for warehouse robots
Tutor Intelligence closed a $75 million Series B led by a16z, taking total funding to $117 million. The company builds a control model for warehouse robots, deployed under the names Sonny and Cassie. It says close to a third of its customers from the past year are returning.
What to act on today
If you are running agents against a long codebase and paying frontier output rates, LongCat-2.5 is worth a priced comparison this week rather than next month — the discount is explicitly limited-time, and the Anthropic-format compatibility means the test costs you a base URL change. If you are evaluating QA automation, read QA Bench for the harness differences rather than the ranking: the same agent with Playwright MCP and with agent-browser produced materially different recall at different cost.
Our running comparison of the testing agents named here is at Best AI QA Agents 2026, with the subscription pricing each one bills against.