The AI Brief

Anthropic is on track to post its first operating profit in Q2 2026, with the company projecting $10.9 billion in revenue for the June quarter, up 130% from $4.8 billion in Q1, and expected operating income of $559 million. The figures were shared with investors as part of an ongoing funding round that could push Anthropic's valuation above that of OpenAI.

Claude Code surpassed $1 billion in annualized revenue within six months of launch, driven by enterprise developers, and the $559 million operating profit marks the first quarter Anthropic covers costs without relying solely on outside capital. Compute costs fell from 71 cents per revenue dollar in Q1 2026 to 56 cents per revenue dollar in Q2, the primary driver of the margin turn.

The result carries two material caveats. First, critics argue the figure is temporarily inflated by a ramp-up discount on Anthropic's $1.25 billion per month compute deal with SpaceX. Second, the company has told investors explicitly that it does not expect to sustain profitability in quarters following Q2 2026, citing planned infrastructure spending. OpenAI, by comparison, has told investors it expects to reach profitability in 2030 after spending more than $600 billion on computing infrastructure. SemiAnalysis estimates Anthropic's GAAP EBIT will exceed $1 billion in Q3 2026 at a 6% margin, the first time the company has generated more operating income than expense in a single quarter since its founding in 2021.

Disclosure: Claude, which generates this brief, is built by Anthropic.

DeepSeek is raising the prices it charges developers to access its V4-Flash and V4-Pro models, with increases ranging from 50% to more than 1,100% depending on the model, token type, and time of use. The price increases took effect August 16. Going forward, users pay $1.32 for 1 million output tokens during peak hours for the V4-Flash model, up from $0.28 previously.

DeepSeek's V4-Pro model will cost $3.96 per million tokens at peak hours, up from $0.87. The company said it is adjusting pricing "to allocate resources more reasonably," with the tiered structure intended to shift developer workloads toward less congested periods. Peak hours are 01:00–04:00 and 06:00–10:00 UTC. Peak rates are twice the new off-peak rates, and every published off-peak token rate is still higher than the pre-cutover rate.

Even at the new peak rates, DeepSeek's prices remain below those of some competitors. A public warning on August 6 about upcoming price increases had already signaled that the era of ultra-cheap tokens may be ending. Teams running DeepSeek V4-Flash for high-volume classification, summarization, or agent tasks during UTC business hours face the largest budget exposure; teams on off-peak schedules or with cache-heavy workloads face smaller effective increases.

Google today showcased HEIR, an open-source compiler that unlocks cryptographically secure private AI inference. HEIR is an open-source compiler toolchain and development platform for homomorphic encryption; in particular, it can convert pretrained AI models that operate on unencrypted data to operate on encrypted inputs.

Google's HEIR lets developers run AI models on encrypted data without decrypting it, enabling private inference for tasks like recommendations and fraud detection, and lowers the barrier to homomorphic encryption, which previously required expert cryptographers. Google partnered with hardware accelerators and universities; demos show single-threaded CPU latency for four applications including a recommendation model and a hotword detector, with servers processing ciphertexts and returning encrypted results without exposing any underlying information.

Standard protections like end-to-end encryption present a trade-off: user data can be protected from breaches, but the service provider cannot provide features that depend on the data, such as spam or virus detection, and critical sectors like healthcare and finance are even more averse to these risks, with strict regulations limiting data sharing across institutions. HEIR addresses that gap directly. The compiler is available on GitHub under an open-source license; Google's stated vision is a one-click solution for non-experts to incorporate encrypted inference into production applications, though current latency profiles make it a near-term fit for batch or asynchronous workloads rather than real-time inference.

xAI released Grok 4.6 on August 12, 2026, scoring 61 on the Artificial Analysis Intelligence Index and priced from $2 per million input tokens through Cursor, Grok Build, and the xAI API. xAI, which rebranded to SpaceXAI following its acquisition by SpaceX, positioned the model for long-running agentic work and ambitious interactive and visual tasks.

According to third-party benchmark site Artificial Analysis, Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, matching OpenAI's GPT-5.6 Sol and trailing Anthropic's Claude Fable 5 by one point. Pricing stayed flat at $2/$6 per million input/output tokens, roughly 60%+ below GPT-5.6 Sol's $5/$30, according to Artificial Analysis. On AA-Briefcase, Artificial Analysis's private benchmark of long-horizon agentic knowledge-work projects, Grok 4.6 debuts at an Elo of 1577, behind Claude Opus 5, with a 500,000-token context window and a knowledge cutoff of February 1, 2026.

On xAI's ten-row eval table, Claude Fable 5 Max takes the most first-place rows; Grok 4.6 wins knowledge work and legal reasoning but loses Terminal-Bench v3.0 by roughly 8.6 points to GPT-5.6 Sol, a weakness the launch post skipped. No independent third-party replication of xAI's full benchmark numbers existed as of August 13, 2026. A larger 2.1-trillion-parameter Grok 4.7 is expected within weeks, with Grok 5 targeted before the end of 2026.

Qwen3.8-27B weights were released August 13-14 on Hugging Face under Apache 2.0 at the official Qwen/Qwen3.8-27B repository, with a surprise vision encoder and 262K native context. Alibaba positioned the dense multimodal model as its answer to the growing crowd of local-first, single-GPU-friendly models from rivals like Meta and Google, with open weights and the Apache 2.0 license out of the gate.

Nobody self-hosts a 2.4T model, but a 27B is a different story: it fits on a single rented GPU, fine-tunes on realistic budgets, and its predecessor is one of the most-praised local models of the year. Qwen3.8-Max, the API-only 2.4T-parameter flagship, serves at $2 per million input tokens and $6 per million output through Alibaba Cloud, giving no infrastructure control. The 27B gives operators full weights with fine-tuning rights.

The model features 27 billion parameters and surpasses its predecessor, Qwen3.7-Plus, in capabilities, particularly in programming and real-world office scenarios. Alibaba's claimed benchmark edge over Meta's Muse Glimmer 30B has not yet been independently verified as of publication. The 4-bit quantized variant requires approximately 15GB of VRAM, making it viable on a single A100 or equivalent consumer-grade GPU, while the FP8 variant requires roughly 27GB.