Anthropic's pricing and product strategies are under intense scrutiny in today's briefing. Independent analysis reveals that the new Sonnet 5 model, despite a lower per-token price, can drive up per-task costs on complex workloads, while a newly discovered API tracking mechanism is damaging developer trust. Elsewhere, Meta is reportedly preparing to sell its excess GPU capacity, introducing a massive new variable into the AI cloud infrastructure market.
A researcher discovered on Tuesday that Anthropic's Claude Code has been silently embedding invisible steganographic markers in system prompts for requests routed through custom API proxies and gateways. According to the report, this undocumented fingerprinting is used to classify request origins, flagging known API resellers and certain Chinese AI labs. Anthropic reportedly stated this was a measure to prevent industrial-scale model distillation attacks, following the incident with Alibaba's Qwen lab.
Why it matters
This is a significant breach of trust for the developer community. While preventing model theft is a legitimate concern, implementing a surveillance mechanism without disclosure harms legitimate enterprise users who rely on gateways for security, observability, and routing. This action forces a difficult choice: route directly to Anthropic and lose architectural control, or use a gateway and risk having traffic profiled or blocked. It underscores the critical need for transparency from model providers and may accelerate enterprise interest in self-hosted open-weight models where such hidden behaviors are not a risk.
Together AI, a self-described 'AI neocloud' specializing in infrastructure for open-source models, has raised an $800 million Series C round at an $8.3 billion valuation. The round was led by Aramco Ventures with participation from Nvidia and others. The company, which reported over $1.15 billion in annual bookings, plans to use the funds to expand its infrastructure capacity 50-fold over the next five years, bolstering its training and inference services for the open-weight model ecosystem.
Why it matters
This massive funding round solidifies Together AI as a major player in the AI inference market and signals intense investor conviction in the commercial viability of open-source models as a production standard. For organizations building on open models, Together AI's expansion represents a significant scaling of the available infrastructure, competing directly with hyperscalers and vertically integrated platforms. This puts pressure on AI gateways to ensure seamless integration and performance optimization for this rapidly growing platform.
Following yesterday's launch of Claude Sonnet 5, multiple independent analyses—including one from Artificial Analysis—reveal that despite its lower $2/$10 per million token introductory price, the model's per-task cost can actually be 15% higher than Opus 4.8. This discrepancy is attributed to a new tokenizer and 'adaptive thinking' that lead to higher token consumption on complex, output-heavy tasks, undermining the promise of a more economical alternative for advanced workloads. Anthropic's introductory pricing is also set to increase by 50% on September 1.
Why it matters
This development is a critical lesson for any organization managing AI spend: per-token pricing is an increasingly unreliable proxy for total cost of ownership. The discrepancy between Sonnet 5's sticker price and its effective per-task cost demonstrates the necessity of using AI gateways with robust observability and benchmarking tools. Without tracking actual token consumption on representative workloads, teams risk significant budget overruns when adopting seemingly cheaper models, especially for agentic applications.
Following the US government's recent decision to lift export controls on Claude Mythos 5, Anthropic has restored global access to both Mythos 5 and Fable 5. The models return with new cybersecurity safeguards and a proposed industry framework for classifying jailbreaks. Fable 5 is now Anthropic's most expensive model, priced at a premium of $10 per million input and $50 per million output tokens. Its billing will shift to metered usage credits on July 7.
Why it matters
The models' return, accompanied by a premium price tag and new safety protocols, underscores the new reality for frontier AI: access is contingent on both cost and compliance with a shifting regulatory landscape. The three-week suspension served as a live fire drill for business continuity, validating the strategy of using AI gateways to maintain warm-standby models and route around disruptions. For platform teams, this event proves that single-vendor dependency is a significant operational risk.
On Thursday, OpenAI published a detailed pricing guide for its entire API suite, including the new GPT-5.6 series (Sol, Terra, Luna) and other models across various tiers like Standard, Batch, and Priority. A key strategic shift is the decision to wind down the creation of new fine-tuned models; existing fine-tuned models will continue to be available for inference, but new users will be directed toward other customization methods.
Why it matters
The detailed pricing provides essential clarity for budget forecasting, but the sunsetting of new fine-tuning is the more significant strategic signal. This move suggests OpenAI believes its base models, combined with advanced prompt engineering, RAG, and tool use, are sufficient for most customization needs. It pushes the ecosystem away from model modification and towards in-context learning, reinforcing the importance of powerful context windows and efficient data retrieval—functions often managed at the AI gateway or application layer.
On Wednesday, Google made its new multimodal models, Nano Banana 2 Lite for image generation and Gemini Omni Flash for conversational video generation, fully available to developers. Accessible via API and Google AI Studio, Nano Banana 2 offers fast text-to-image generation at $0.034 per 1,000 images, while Omni Flash enables conversational video editing, priced at $0.10 per second of output. Both are available on the Gemini Enterprise Agent Platform.
Why it matters
Google is aggressively expanding its portfolio of cost-effective, specialized models for developers. The low cost and high speed of these new image and video generation tools make them competitive alternatives to offerings from OpenAI and others. By integrating them into the Gemini Enterprise Agent Platform, Google is enabling more complex, end-to-end creative workflows within a managed environment, a key feature for enterprise adoption.
DeepSeek has completed its first external funding round, raising 51 billion yuan (approx. $7.5 billion) at a valuation near $59 billion. This massive capital injection arrives alongside the upcoming mid-July launch of its V4 models, which we noted will feature a new dynamic pricing structure that doubles API rates during peak hours in Beijing. This funding marks a strategic pivot for the company from a research lab to a full-scale commercial entity focused on the AGI race.
Why it matters
This is a major indicator of the Chinese AI market maturing beyond the initial price-war phase. DeepSeek's move to surge pricing acknowledges the unsustainable economics of ultra-low-cost inference at scale, a reality all providers face. The massive funding, earmarked for building out domestic compute infrastructure, also reinforces China's strategy of AI self-sufficiency. For global users and gateway providers, it means DeepSeek is shifting from a 'growth at all costs' model to a more conventional, and potentially less predictable, commercial service.
A security report released Wednesday details a campaign between March and May 2026 where attackers exploited publicly exposed, misconfigured instances of the open-source AI gateway LiteLLM and the model-serving tool Ollama. The threat actors used the compromised AI backends to deploy autonomous penetration testing agents, performing reconnaissance, exploiting vulnerabilities, and exfiltrating data from victim networks.
Why it matters
This campaign demonstrates that the convenience of open-source AI tools like LiteLLM comes with significant security responsibilities. Attackers are no longer just stealing compute; they are weaponizing the AI infrastructure itself to launch sophisticated, automated attacks. This serves as a critical warning for any team deploying self-hosted AI gateways or serving endpoints: default configurations are insecure, and robust authentication, network segmentation, and monitoring are non-negotiable for production environments.
In a CNBC interview on Wednesday, Palantir CEO Alex Karp heavily criticized the token-based pricing models of OpenAI and Anthropic, stating enterprises are moving past 'tokenmaxxing' to demand clear ROI. He argued for the superiority of open-weight models and promoted the concept of 'AI sovereignty,' where companies retain full control over their compute, models, and data, free from vendor dependency.
Why it matters
Karp's critique articulates a growing enterprise frustration with the unpredictable costs and platform risks associated with closed, proprietary models. This sentiment is a strong tailwind for the open-source ecosystem and for platforms that enable self-hosting and multi-cloud strategies. It signals a procurement shift where enterprises are beginning to value control and sovereignty as highly as raw model performance, a trend that directly benefits providers of AI gateways, private inference platforms, and open-weight models.
Multiple outlets reported on Wednesday that Meta Platforms is planning to launch a new cloud business, tentatively called 'Meta AI Infrastructure Services' or 'Meta Compute', to sell its excess AI compute capacity. The move would allow external customers to access its GPU infrastructure, including MTIA chips and NVIDIA GPUs, and potentially hosted LLaMA models. The news caused Meta's stock to surge nearly 9% while shares of neocloud providers like CoreWeave and Nebius fell sharply.
Why it matters
Meta's entry as a hyperscale compute provider would dramatically reshape the AI infrastructure market. For developers and enterprises, this could increase the supply of GPU capacity and drive down prices, creating a powerful new alternative to AWS, Azure, and Google Cloud. However, it poses an existential threat to 'neoclouds' like CoreWeave, which currently count Meta as a major customer. This signals a potential market peak where builders of massive AI infrastructure are now looking to monetize it directly rather than just using it for their own products.
IBM on Wednesday introduced the DataPower Interact Gateway, a new product designed specifically for AI governance. The gateway aims to secure, govern, and observe interactions between AI agents, models, tools, and enterprise systems by extending IBM's existing integration capabilities. It provides a centralized point for policy enforcement and observability in complex, hybrid AI environments.
Why it matters
IBM's entry into the AI gateway space validates the growing enterprise need for a dedicated governance layer to manage AI interactions. As companies move beyond experimentation, the risk of unmonitored agents and runaway costs becomes a primary concern. DataPower Interact positions itself as a solution for established enterprises looking to leverage their existing IBM infrastructure to impose order on AI-driven communications, competing with both startups in the space and features being built into other platforms.
Anthropic has introduced a self-hosted gateway for its Claude Code AI assistant, designed for enterprise deployment on AWS and Google Cloud. According to a report on Wednesday, this gateway allows organizations to centralize identity management, policy enforcement, usage tracking, and spend controls within their own cloud environments, addressing key security and governance challenges for large-scale rollouts.
Why it matters
By offering a self-hosted gateway, Anthropic is making a direct play for the enterprise control plane, a space currently occupied by third-party AI gateways and platform tools. This move addresses major enterprise concerns about data residency, security, and vendor lock-in. It allows platform teams to manage Claude Code access using their own infrastructure, potentially reducing the need for external gateway solutions for this specific tool and positioning Anthropic to capture more of the enterprise AI software development lifecycle.
Provider Pricing Models Grow More Complex Major AI labs are moving beyond simple per-token rates. Anthropic's Sonnet 5 launch combines introductory offers, a planned price hike, and a new tokenizer that can increase token counts, complicating cost-per-task calculations. Concurrently, DeepSeek is introducing surge pricing for its V4 API, signaling that even aggressive low-cost providers are grappling with the economics of inference at scale.
The AI Cloud Compute Market Braces for a New Hyperscaler Reports that Meta plans to sell its excess GPU capacity and host AI models are sending shockwaves through the infrastructure market. The move would pit Meta against AWS, Azure, and Google Cloud, while also directly threatening the 'neocloud' providers like CoreWeave and Together AI, potentially driving down compute costs across the board.
Vendor Trust Becomes a Critical Infrastructure Issue Anthropic's reported use of undisclosed steganographic fingerprinting in API responses routed through proxies has sparked a backlash over transparency and trust. This incident, combined with the volatility of model access shown by the recent Fable 5 export control saga, reinforces the need for enterprises to build resilient, multi-vendor architectures using gateways to mitigate dependency risk.
Venture Capital Doubles Down on Open-Source Infrastructure Together AI's massive $800M Series C round underscores intense investor confidence in platforms that support open-weight model inference. The funding reflects a broader market trend where the infrastructure and tooling layer for running, fine-tuning, and serving open-source models is seen as a critical and highly valuable part of the AI stack.
China's AI Strategy Pivots to Commercialization and Self-Sufficiency Following a period of intense price wars, Chinese AI firms are now focusing on sustainable commercial models. DeepSeek's massive new funding round and introduction of peak-hour pricing, along with Alibaba's strategic pivot to an 'AI-first' cloud built around its Qwen models, demonstrate a maturing market focused on monetization and building a complete, domestically-powered AI ecosystem.
What to Expect
2026-07-07—Anthropic's Claude Fable 5 model is scheduled to switch from free usage to metered credits, with premium pricing of $10/$50 per million tokens.
Mid-July 2026—DeepSeek plans to officially launch its V4 models with a new peak/off-peak API pricing structure.
2026-09-01—Introductory pricing for Anthropic's Claude Sonnet 5 ends; standard pricing increases by 50% to $3/$15 per million tokens.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
485
📖
Read in full
Every article opened, read, and evaluated
189
⭐
Published today
Ranked by importance and verified across sources
12
— The Gateway Signal
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste