GLM-5.1 Tops GPT-5.4 as SAP Closes Agent APIs to Third Parties

Two Moves, One Pattern: The Data Layer Is the New Moat

A product manager at a mid-market ERP vendor spent Tuesday morning reading two news items and closed the tab with a worse roadmap than she woke up with. SAP blocked all third-party AI agents from its APIs, reserving access for its own Joule and Nvidia's NemoClaw, and in the same week committed €1B to acquire Prior Labs for a proprietary lab focused on structured enterprise data. The same week, Anthropic shipped 10 purpose-built finance agents covering pitchbooks, credit memos, KYC, and month-end close, wired through Microsoft 365 and Moody's data. FactSet dropped 8% in a session.

Read together, the two moves describe one pattern. SAP is pulling the fence inward. Anthropic is routing around the fence by partnering with alternative data (Moody's) and reaching the workflow through Microsoft's pipes. The thing being pitched is "AI agents." The thing actually being done is a renegotiation of who charges rent on enterprise data access.

The question on a PM's desk is no longer 'how good is our AI' — it is 'who owns the pipes our AI needs, and have they started charging rent yet.'

Sierra Proves the Standalone Agent Category Is Real

Sierra crossed $150M ARR serving more than 40% of the Fortune 50, and raised $950M at a $15B+ valuation. Those are production deployments at companies whose procurement takes two quarters to say yes. Bret Taylor separated the thing most teams conflate: an "AI assistant for X" is not the same product as "AI that owns the outcome of X." Enterprise buyers paid separately, rather than wait for the incumbent's marketplace, because the second framing commands roughly 10x the monetization.

Short Sellers Confirm the Displacement Clock

Viceroy Research, the firm that shorted Wirecard, is now publicly screening for "high margin businesses with clean balance sheets that happen to be in the crosshairs of a new technology roadmap." That is professional capital running the displacement thesis as a systematic filter. Competitive narrative has to answer the AI-moat question before the short report lands, not after.

The 2x2 for Your Integration Roadmap

Any integration roadmap that assumes a major platform's API will stay open should be rewritten this sprint, with longer procurement cycles and rev-share priced in. The work this week is naming the one integration on the roadmap to kill before engineering spends another six weeks on an API that may be gone by the time the feature ships.

The Numbers That Changed This Week

A platform lead pulled up her inference contract this week and noticed the renewal is in February. Zhipu AI released GLM-5.1, a 744B-parameter MoE model (40B active per token) under an MIT license. It scores 58.4 on SWE-Bench Pro, above GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3. It ships with 200K context and 8-hour autonomous execution. This is not "almost as good." It is 0.7 points better than the closed frontier on a coding benchmark, and the license costs nothing.

The same week, xAI priced Grok 4.3 at $1.25 per million input tokens ($2.50 output) with 1M context, always-on reasoning, native multimodal, and live web search. That is roughly 5x cheaper than GPT-5.4 and Claude Opus 4.7 for equivalent-class capabilities.

Any unit economics model that assumed the inference cost of six months ago is now either leaving margin on the table or exposed to a competitor who will pass the savings through.

The Sentiment Shift Has Numbers

A poll of 201 AI practitioners found 45% believe OpenAI has lost its default-lab position. Only 16% see it as a two-horse race with Anthropic. Another 20% think open-weight models close the gap first. That is sentiment, not usage. Practitioner sentiment leads procurement behavior by 6-12 months. OpenAI meanwhile is redirecting attention to government distribution: $1 ChatGPT for federal agencies, a $200M Pentagon contract, lobbying spend up 577% to $1.76M. A lab winning on model quality does not need to price ChatGPT at a dollar.

Separate the Two Decisions

Teams keep bundling these into one story. They are two different decisions with two different owners:

1. The open-weights decision: whether to self-host GLM-5.1 for coding and structured workloads. This is an infra investment with a capex payback calculation, and it belongs to the platform team.
2. The routing decision: whether to send cheap or tolerant traffic to Grok 4.3 while keeping the incumbent on quality-critical paths. This is a sprint-level change, not a migration, and it belongs to whoever owns the gateway.

Google DeepMind's Gemma 4 Multi-Token Prediction drafters add a third lever: 3x inference speedup without quality loss via speculative decoding, available across vLLM, MLX, and Transformers. For a voice agent or a code-completion surface where the 2-3 second round trip was the real product blocker, sub-second becomes achievable.

The Decision Framework

The honest recommendation: staff one cell this sprint and let the other wait a sprint. Hedging both reads prudent in the deck and costs a quarter on the roadmap. Use Grok 4.3 at $1.25/M as the contract renewal anchor regardless of which cell gets the headcount.