Salesforce agrees to buy Braze
Salesforce said on 20 October that it will acquire Braze, the NASDAQ-listed customer engagement platform, for $47.50 in cash per Class A share and about $5.66 billion in enterprise value (Exhibit 99.1 to Salesforce's Form 8-K). Closing is expected in Salesforce's fiscal year 2028, subject to customary conditions including a Braze shareholder vote and regulatory approvals. Salesforce said it anticipates no change to FY27 guidance, has not guided FY28, and has no update to its long-term targets.
Braze's real-time engagement and AI decisioning are expected to complement all four layers of Salesforce's architecture, Agentforce among them, and BrazeAI adds optimisation across content, offers, channels and timing. European teams running Braze should ask, before close, which entity will process their customer event data and which sub-processor list covers it.
Artificial Analysis publishes HealthBench scores, and reply length changes the order
Artificial Analysis published leaderboards on 19 October for HealthBench and HealthBench Professional, two OpenAI benchmarks graded against physician-created rubrics: 5,000 conversations with everyday users and health workers in the first, 525 conversations all with clinicians in the second. The length-adjusted score lowers the score of replies longer than 2,000 characters and raises shorter ones. Claude Opus 5.5 holds the highest raw score on HealthBench Professional at 71.4% and drops to sixth once adjusted, at 59.9%. GPT-6 Astra tops the adjusted table at 66.0%, comparable to GPT-6.1 Sol at 65.1% and ahead of Claude Sonnet 5.5 at 63.0%.
Opus 5.5 and Claude Fable 5.1 average about 5,900 characters per answer on the clinician set, nearly three times the threshold. GPT-6 Sol is the only model under it, at about 1,600 characters, and on the larger HealthBench it writes about 1,000, where it has the highest raw communication-quality score at 81% and the lowest completeness score at 34%. The quoted figures are each model's maximum configuration, the Claude entries max with fallback. A vendor quoting a HealthBench figure should be asked which of the two scores it is.
Quick Hits
- MiniMax's Flash model has a score before it has weights. — Artificial Analysis puts MiniMax-M3.1-Flash at 43 on its Intelligence Index and lists the weights as not public. The Decoder reports MiniMax plans to open them on 23 October.
- Oracle and AstraZeneca signed a non-binding AI agreement. — The 20 October agreement says initial areas could include oncology, cardiovascular, renal and metabolic disease, respiratory and immunology, and rare disease, with clinical-trial recruitment among the aims. No value was disclosed.
- Microsoft pulled out of a 50MW Nscale site in Essex. — Grid officials told Nscale they did not expect to power the Loughton site (23,040 Nvidia GB300 GPUs for Azure) until the 2030s, Reuters reported. Nscale is exploring alternative power sources.
- European tech funding ran past €1.8 billion last week. — Tech.eu counted more than 65 deals, among them a $50M Series B for UK AI data-security firm Harmonic Security.