Fable 5 Restored After a 19-Day Outage, Sonnet 5 Launches at $2/$10 Promotional Pricing, and OpenAI’s First In-House Inference Chip Debuts — Government Directives Become a Model Availability Variable

Major Developments

A 19-Day Outage Triggered by One Directive: The Fable 5 Export Control Episode and Restoration

On June 12, the U.S. government applied export controls to Anthropic’s latest Claude Fable 5 and Mythos 5 models, requiring access restrictions for all foreign nationals. Because the directive took effect immediately and Anthropic had no real-time mechanism for verifying users’ nationalities, the company chose to suspend access to both models for all users. The trigger was a prompt technique identified by Amazon researchers that bypassed Fable 5’s safeguards, allowing the model to identify several software vulnerabilities and, in one case, generate demonstration code showing how a vulnerability could be exploited. Anthropic later said its testing showed that several less capable models — including Opus 4.8, GPT-5.5, and Kimi K2.7 — could identify the same vulnerabilities, and that all tested models could generate the same exploit demonstration. In other words, the jailbreak did not unlock a unique Mythos-level capability. On June 26, the government approved the restoration of Mythos 5 access for a limited set of U.S. organizations; on June 30, the Department of Commerce lifted the controls; and on July 1, Fable 5 was restored globally. For Pro, Max, Team, and select Enterprise plans, Fable 5 is available for up to 50% of weekly usage limits until July 7. Channel restoration is staggered: first-party platforms resumed access on the day of restoration, while access through AWS, Google Cloud, and Microsoft Foundry will follow “as quickly as possible” according to Anthropic. Cloud distribution channels are therefore recovering more slowly than direct access through the model provider.

On the technical mitigation side, Anthropic worked with the government to train a new safety classifier targeting the jailbreak technique, with a reported block rate of over 99%. Blocked requests are redirected to Opus 4.8, at the cost of a higher false-positive rate for ordinary coding and debugging requests. Anthropic also began drafting an industry framework for scoring jailbreak severity together with Glasswing partners including Amazon, Microsoft, and Google, and committed to providing the government with pre-release evaluation access for frontier models. Viewed against the background of the June 2 White House executive order, these measures suggest that the release and operation of U.S. frontier models are becoming institutionally embedded in government processes.

The direct lesson for enterprises is that the availability risk register now has a new category: a regulatory directive can cut off global access to a model within hours, with no advance notice and no transition period, and restoration timing depends on public-private negotiation rather than vendor SLA. During the 19-day outage, teams that had bound critical workflows to a single model had little choice but to wait. By contrast, customers that had configured cross-vendor fallback chains on AgentsFlare could automatically shift traffic by policy to Opus 4.8, GPT-5.5, or domestic open-weight models. The audit trail fully recorded the switching process, reducing the burden of post-incident compliance reconstruction. As geopolitical and regulatory variables at the model layer increase, fallback capability is moving from architectural neatness to practical insurance.

Sonnet 5 Launches: A One-Third Nominal Price Cut That Is Roughly Cost-Neutral After Tokenization

On June 30, Anthropic released Claude Sonnet 5, positioning it as its most capable agentic Sonnet model to date. Anthropic’s official benchmarks show reasoning, tool use, and coding performance approaching Opus 4.8, with a default context window of 1 million tokens. Pricing adopts an introductory structure: through August 31, Sonnet 5 is priced at $2 per million input tokens and $10 per million output tokens; from September 1, it returns to standard pricing of $3/$15, in line with the previous Sonnet 4.6 price. For reference, Opus 4.8 is priced at $5/$25.

Outside the price table, there is one easily overlooked change in measurement basis: Sonnet 5 uses a new tokenizer. The same input can map to 1.0–1.35× as many tokens, depending on content type. Anthropic itself notes in the launch materials that the introductory pricing is designed to make the transition roughly cost-neutral during the migration period. In other words, moving from Sonnet 4.6 to Sonnet 5 during the promotional period should result in broadly similar bills; after standard pricing resumes on September 1, the same workload could cost up to 35% more than under Sonnet 4.6. Enterprises should not simply multiply historical token usage by the new list price when budgeting. They need to test the conversion ratio using their own real workloads and separately record token metrics before and after the model switch in cost attribution. AgentsFlare’s model-level and team-level cost records can directly support this kind of before-and-after comparison.

The practical impact on model selection lies in the tier structure. Sonnet 5’s introductory price is only 40% of Opus 4.8’s, while the performance gap between the two has narrowed across many agentic tasks. With adjustable effort levels, even the same model can represent several multiples of cost variation. For agentic workloads, moving the default model down from the flagship tier to Sonnet 5 and reserving Opus for difficult tasks is the most direct cost-reduction lever this quarter.

OpenAI’s First In-House Chip, Jalapeño: A Nine-Month Design Cycle and Year-End Deployment

On June 24, OpenAI and Broadcom jointly unveiled Jalapeño, OpenAI’s first in-house chip. It is positioned as a dedicated LLM inference ASIC — a purpose-built chip for a single category of workload — rather than a training chip or general-purpose accelerator. The chip is close to the reticle-size limit for a single lithography exposure. It took only nine months from project launch to completion of design, and OpenAI said its own models played a substantial role in the design process. Celestica is responsible for boards and rack systems, while networking uses Broadcom’s Tomahawk networking silicon. Deployment is planned to begin by the end of 2026. OpenAI’s official framing is that the chip delivers substantially better performance per watt than the current state of the art and is designed to work with LLMs across the industry, not only GPT models.

This chip does not change market supply in the short term. It is not sold externally, and capacity will be prioritized for OpenAI’s own use. The medium-term transmission path is on the cost side: if a meaningful portion of OpenAI’s inference cost structure shifts from purchasing NVIDIA systems to amortizing its own ASICs, its room to maneuver on API pricing expands. For OpenAI, which is in an intense pre-IPO pricing and valuation negotiation window, this is real strategic leverage. For NVIDIA, this is another signal — following Google TPU and Amazon Trainium — that a major customer is shifting inference workloads toward custom silicon. The training market remains relatively secure for now, but pricing power in inference is slowly being eroded.

Qualcomm Drops Tenstorrent and Buys Modular: $3.9 Billion for the Software Layer

Last week’s brief tracked reports that Qualcomm was discussing an $8–10 billion acquisition of Tenstorrent. This week brought a reversal: on June 30, Tenstorrent CEO Jim Keller publicly denied in Tokyo that the company was in acquisition talks with Qualcomm, saying that Tenstorrent remains focused on its own business. Earlier, on June 24, Qualcomm had confirmed an approximately $3.9 billion all-stock acquisition of Modular, with the transaction expected to close in the second half of 2026. Modular was founded by former Apple and Google engineer Chris Lattner. Its product is a cross-hardware inference software stack that allows the same model code to run on accelerators from different vendors without rewriting for each hardware target. It is widely viewed as one of the most developed portability layers outside CUDA.

Taken together, the two developments clarify Qualcomm’s path into the data center: build the chips in-house and acquire the missing software ecosystem. The lesson for enterprises is the direction this deal confirms. Inference hardware is diversifying — NVIDIA, AMD, custom ASICs, and domestic chips — and the player that provides a unified software layer across hardware will control the entry point into this fragmented landscape. As an independent company, Modular occupied a neutral-layer position. After being brought into Qualcomm, that neutrality is less certain, making the gap at this layer even more visible.

LiteLLM Gateway Vulnerability Chain Exploited in the Wild

On June 8, CISA added LiteLLM command injection vulnerability CVE-2026-42271 (CVSS 8.7) to its Known Exploited Vulnerabilities Catalog, confirming exploitation in the wild. The vulnerability sits in the MCP server’s test endpoint. A user with any proxy API key can execute arbitrary commands on the gateway host. When chained with the Starlette framework’s Host header validation bypass vulnerability (CVE-2026-48710), an attacker can remotely execute code without any credentials; the combined chain carries a CVSS score of 10.0. Compromising a LiteLLM gateway means gaining access to all model-provider API keys stored there and potentially moving laterally into downstream systems. Fixed versions are LiteLLM 1.83.7 and Starlette 1.0.1. Mitigations include blocking the two test endpoints at the reverse proxy layer and rotating all credentials stored in the gateway. Self-hosted AI gateways centrally store an enterprise’s most sensitive model credentials and have become high-value targets for attackers. Patch timeliness for these components should now be part of the routine security operations checklist.

Upstream Compute: Diverging Trends in GPU Rental and Memory

GPU rental prices are falling. Industry rental price data shows NVIDIA B200 hourly rental prices falling from around $6.11 to $4.22 within three weeks, a drop of roughly 30%. Drivers include improved yields on TSMC’s 4NP process, which has reduced system-level costs, easing HBM3e supply, and competition from emerging clouds such as RunPod, Lambda, and Nebius after they obtained spot inventory. Falling rental prices mean inference compute availability is improving, easing short-term cost pressure for small and medium-scale deployments.

Memory is moving in the opposite direction. TrendForce’s June 30 DRAM contract price update continued to show rising prices, while the market expects conventional memory output to contract. The firm’s early-June analysis projected that HBM contract prices — HBM being high-bandwidth memory formed by vertically stacking multiple memory chips and packaging them closely with the GPU — could rise by multiples in 2027, with negotiations between buyers and sellers already shifting toward 2027 HBM4 supply agreements. Micron had previously disclosed that its 2026 HBM capacity was largely sold out. At the market-signal level, capital is rotating from the GPU segment toward memory. The meaning of these opposing trends is that rental-market price reductions are being driven by supply-side competition and one-off improvements in system cost, while the structural shortage in upstream memory has not eased. Memory continues to account for a rising share of AI server bill of materials, and next year’s cost curve for new compute capacity will remain constrained by HBM rather than the GPU die itself.

A supporting capacity signal came from Blackstone President Jonathan Gray, who confirmed in a Nikkei interview this week that Blackstone plans to invest $30 billion over the next three to five years in AI data centers in Japan. It has already built more than 500MW and is discussing 1GW-scale projects. Inference capacity supply in the Asia-Pacific region is accelerating, and enterprises with deployment or data residency needs in Japan and surrounding markets will have more options.

Integrated Assessment: Regulatory Switches, Pricing Illusions, and the July Release Window

First, government action has formally become a model availability variable. The 19-day Fable 5 outage is the first global case: the trigger was a jailbreak report, the execution was an immediate full shutdown, and restoration depended on public-private negotiation. The jailbreak severity framework, pre-release government evaluations, and the vulnerability information-sharing mechanism established by the executive order will likely make future frontier model releases more orderly. But they also mean that access to models from U.S. providers is shifting from a purely commercial matter to one with government process embedded in it. Shutdown and restoration timelines are no longer fully controlled by vendors. When enterprises assess suppliers, the regulatory jurisdiction in which a model sits is now a practical parameter alongside price, performance, and SLA.

Second, model-layer price sheets are becoming less informative. With Sonnet 5’s introductory price, tokenizer conversion, and effort levels layered together, the actual unit cost of the same model can vary by several multiples. Cross-model price comparisons require a normalized basis before they are meaningful. The downward trend in nominal unit prices is real, but the actual cost enterprises receive depends on workload structure and configuration details. The error margin of budgeting from list prices will continue to widen. Testing against real traffic is the only reliable comparison method.

Third, the July release window is crowded. After missing the June 30 target window, Google shifted Gemini 3.5 Pro to a July launch. Multiple media reports indicate that broad rollout of GPT-5.6 is expected in mid-to-late July. Together with the restoration of Fable 5 and Sonnet 5’s promotional period, the first month of the third quarter is likely to reshuffle model options, pricing, and capability rankings again. Locking into a long-term single-model contract this quarter carries elevated timing risk.

References

Anthropic — Redeploying Fable 5 (outage timeline, jailbreak report, industry framework): https://www.anthropic.com/news/redeploying-fable-5 — 2026-06-30

CNBC — Anthropic says Trump admin has lifted export controls: https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html — 2026-06-30

Anthropic — Introducing Claude Sonnet 5 (pricing, tokenizer, benchmarks): https://www.anthropic.com/news/claude-sonnet-5 — 2026-06-30

Claude Platform Docs — What's new in Claude Sonnet 5 (1M context): https://platform.claude.com/docs/en/about-claude/models/whats-new-sonnet-5 — 2026-06-30

OpenAI — OpenAI and Broadcom unveil LLM-optimized inference chip: https://openai.com/index/openai-broadcom-jalapeno-inference-chip/ — 2026-06-24

Tom's Hardware — Jalapeño reticle-sized ASIC, nine-month cycle: https://www.tomshardware.com/tech-industry/artificial-intelligence/broadcom-and-openai-unveil-custom-built-jalapeno-inference-processor-openais-first-chip-is-a-massive-reticle-sized-asic-built-in-an-ultra-fast-nine-month-development-cycle — 2026-06-24

Qualcomm — Qualcomm to Acquire Modular: https://www.qualcomm.com/news/releases/2026/06/qualcomm-to-acquire-modular — 2026-06-24

Bloomberg — Qualcomm Confirms Buying Modular (approximately $3.9B all-stock): https://www.bloomberg.com/news/articles/2026-06-24/qualcomm-confirms-buying-modular-to-help-ai-market-push — 2026-06-24

GuruFocus — Tenstorrent CEO Denies Qualcomm Acquisition Talks: https://www.gurufocus.com/news/8938157/tenstorrent-ceo-denies-qualcomm-acquisition-talks-amid-focus-on-ai-development — 2026-06-30

The Hacker News — LiteLLM Flaw CVE-2026-42271 Exploited in the Wild: https://thehackernews.com/2026/06/litellm-flaw-cve-2026-42271-exploited.html — 2026-06-09

Horizon3.ai — CVE-2026-42271 chained with CVE-2026-48710: https://horizon3.ai/attack-research/vulnerabilities/cve-2026-42271-chained-with-cve-2026-48710/ — 2026-06

Thunder Compute — AI GPU Rental Market Trends (July 2026): https://www.thundercompute.com/blog/ai-gpu-rental-market-trends — 2026-07

TrendForce — DRAM Price Trends (6/30 contract price update): https://www.trendforce.com/price/dram/dram_spot — 2026-06-30

TrendForce — HBM Contract Prices Expected to Surge Multiples Higher in 2027: https://www.trendforce.com/presscenter/news/20260602-13074.html — 2026-06-02

Nikkei Asia — Blackstone to invest $30bn in Japan AI data centers: https://asia.nikkei.com/editor-s-picks/interview/blackstone-to-invest-30bn-in-japan-ai-data-centers-president — 2026-07

TechTimes — Gemini 3.5 Pro Cleared for July Launch, GPT-5.6 Stays Locked: https://www.techtimes.com/articles/319318/20260629/gemini-35-pro-cleared-july-launch-fable-5-nears-return-gpt-56-stays-locked.htm — 2026-06-29