Reflection's Beam splits two coding benchmarks with GLM 5.2 before its weights ship

Reflection says Beam matches GLM 5.2 on reasoning while using three to four times less inference compute ^[22]. The estimate draws on Artificial Analysis and DataCurve benchmark data. It excludes prompt prefill, context-dependent attention and serving overhead ^[23]. That leaves per-token model compute. A sparse model looks good on that measure: 23 of 501 billion parameters is about 4.6 percent of the weights per token ^[17]. For the ratio to carry over to a real bill, the excluded terms would have to be small for a team's traffic, or shrink by the same factor. Beam is aimed at coding, reasoning and agentic tasks ^[3]. In long agent sessions, the context-dependent attention term grows with the context.

Memory follows the total count. A server has to keep all 501 billion parameters resident, because the router can send the next token to any expert ^[4]. Keeping them resident is a serving cost, and serving overhead is one of the excluded terms ^[23]. RuntimeWire, which reported the scorecard, called the estimate a model-compute comparison and not a measure of what a customer will pay to run the system ^[23].

The two benchmark gaps point in opposite directions. Beam trails GLM 5.2 by 0.9 points on Terminal-Bench 2.1 ^[7][18]. It leads by 3.4 points on SWE-bench Pro v1 ^[8][19]. Both are company-published comparisons, and the company gives no single overall ranking ^[9]. Reflection printed the rows Beam loses, and I give it credit for that. For a model sold on agentic work, second place on the terminal benchmark by under a point is an awkward spot. Either gap only tells a team something if its own tasks look like the benchmark's.

The training spend behind those scores is large. If every GPU in the reinforcement-learning run worked all 28 days, the run used about 7.06 million GPU-hours ^[20]. Reflection also says it trained Beam on 23.8 trillion tokens ^[5]. Semafor reported that the June funding round closed at a $25 billion pre-money valuation, with Nvidia, Sequoia and Citigroup among the investors ^[13].

According to the Financial Times, open-weight models' share of tokens on Vercel's AI Gateway went from 7 percent in December to 56 percent in August ^[11]. That is an eightfold rise in share, measured on one gateway's traffic ^[21]. The FT also reports partnerships with the Pentagon and the US Department of Energy, plus agreements to build models for US allies including South Korea ^[14]. Laskin co-founded Reflection with Ioannis Antonoglou, another former Google DeepMind researcher ^[1]. He told the FT that Reflection is seeing demand from organizations looking for sovereign AI systems and unwilling to rely on Chinese models ^[15].

I'd keep Beam out of capacity plans until outside teams have run the released weights on their own traffic. In my context, which is coding agents with long prompts, the costs the estimate leaves out are the ones I'd need measured first ^[23]. Waiting costs a few weeks ^[10].