Meta's New Model & The AI Pricing Debate

Meta launched Muse Spark 1.1 this week, its most capable model to date, developed by the Meta Superintelligence Labs under Alexandr Wang. On agentic benchmarks, the model rivals GPT-5.5 and Opus 4.8. Muse Spark supports a 1M-token context window with active context management and is priced at $1.25 input and $4.25 output per million tokens, making it 4-8x cheaper than Anthropic’s Fable 5. Zuckerberg framed the pricing as a response to what he called “very extreme” margins from competing labs.

The timing of Muse Spark 1.1’s launch couldn’t be better. Palantir CEO Alex Karp went viral just last week after telling CNBC that enterprise customers were “livid” with frontier labs, that something has “gone completely wrong” with token-based pricing, and that companies are privately telling him they’re getting “no value” from their AI spend. With this week’s launch, Meta is offering a proprietary frontier-tier model at notably cheaper pricing.

What makes this story especially noteworthy is that until recently, Meta was widely considered an afterthought in the model race, which was characterized by SemiAnalysis as a “two horse race between OpenAI and Anthropic. “ Muse Spark 1.1 is not yet at performance parity with Fable 5 or GPT-5.6 Sol, but it is competitive on agentic tasks and dramatically cheaper. Whether that tradeoff resonates with enterprises currently rethinking their AI budgets is the question that will define whether Meta has actually re-entered the race.

On July 7, just 15 days after its IPO, SpaceX (SPCX) joined the Nasdaq-100, marking the fastest-ever entry for a newly listed company following Nasdaq’s May change to its fast-track rules. JPMorgan has estimated that QQQ-tracking funds drove $4.3 billion in buying around the index inclusion. Despite this mechanical demand, SPCX shares fell nearly 7% on their first day in the index at $143, only modestly above the IPO price and down roughly 34% from the all-time high of $225.64 on June 16.

Investor reaction has been mixed, with banks including BofA, Deutsche Bank, Goldman Sachs, UBS, RBC, Bernstein, Morgan Stanley, and JPMorgan bullish on the stock. Insider lockups for stock sales do not begin unwinding until late July, with nearly 20% after Q2 earnings and around 7% in tranches through October, with Musk’s 6.4 billion shares locked until June 2027.

Separately, SpaceX President Gwynne Shotwell announced that she and her husband would donate shares to over 2 million Trump administration-sponsored accounts for children, worth about $320 million. At the White House opening-bell ceremony celebrating the rollout, Trump was joined by Nasdaq CEO Adena Friedman and president Nelson Griggs, plus NYSE president Lynn Martin, though notably none of the five ETFs (SPYM, IVV, ITOT, VTI, SPTM) these accounts can invest in is a Nasdaq fund.

Two unrelated stories have been circulating within the tech world this week surrounding implicit messages in Claude logs. The first was raised by developer “Thereallo” in late June, who found that Claude Code embedded steganographic markers in the system prompt, subtly altering the date format and apostrophe in “Today’s date is,” to flag users on Chinese time zones, proxies, or AI-lab domains, with the domain list XOR-obfuscated. Anthropic engineer Thariq Shihipar called it a March “experiment” to prevent reseller abuse and distillation, which was later rolled back; in response, Alibaba banned Claude Code, and privacy advocates called it a breach of trust.

Separately, on July 6th, Anthropic published interpretability research introducing the “Jacobian lens,” which surfaces the words a model is poised to say. Collectively, these form the “J-space,” a small privileged set of representations behaving like a global workspace, Claude’s analog of accessible thought. Reading it lets researchers catch Claude silently noticing it’s being tested, weighing manipulation, or reacting emotionally, work that has reignited the machine-consciousness debate.

• China landed a reusable Long March rocket booster on a recovery ship, using shipboard netting instead of the landing legs used by SpaceX’s Falcon 9, making it the second country to master orbital-class booster recovery after the US.

• Meta drew up plans for a new cloud infrastructure business to sell excess AI compute, competing directly with AWS, Azure, and Google Cloud through an internal division called Meta Compute that would let outside developers pay to run queries against models, including Muse Spark; the announcement raised Meta shares 10% while CoreWeave and Nebius fell more than 10% each.

• OpenAI released GPT-Live, a new full-duplex voice-model architecture powering ChatGPT Voice that can listen and speak simultaneously, delegate deeper questions to GPT-5.5 in the background, and show visual widgets for weather, stocks, and sports; the new model becomes the default for Go, Plus, and Pro users, and GPT-Live-1 mini for Free users worldwide, replacing the turn-based Advanced Voice Mode.

• Anthropic signed a 20-year, $19 billion lease with crypto miner-turned data center provider TeraWulf for capacity in Hawesville, Kentucky, covering 401 megawatts that will come online in phases starting in late 2027, exceeding TeraWulf’s $12 billion market cap and sending shares up as much as 19% intraday.

• Beijing convened meetings with Alibaba, ByteDance, and Z.ai to consider restricting overseas access to China’s most advanced AI models, including models not yet released, according to three people familiar with the discussions.

• Illinois Gov. JB Pritzker signed the Artificial Intelligence Safety Measures Act into law, requiring large frontier AI developers (>$500 million annual revenue) to publish catastrophic-risk mitigation plans, submit to first-in-the-nation annual third-party audits, and report incidents within 72 hours.

• The US Cybersecurity and Infrastructure Security Agency deployed Anthropic’s Mythos to audit government software, with CISA’s Attack Surface Evaluation team using the model to scan government code repositories for vulnerabilities exploitable by foreign spies and cybercriminals, reportedly recovering a large number of bugs, even as Anthropic remains formally on the Pentagon’s supply-chain risk list.

• Anthropic launched Claude Science, an AI workbench for scientific research, alongside its own internal drug-discovery program focused on neglected diseases, demonstrated during a livestreamed event screening 2.2K compounds to find a stabilizer for the enzyme responsible for phenylketonuria.

• SpaceXAI and Cursor released Grok 4.5, their first jointly developed AI model built for coding, finance, and legal tasks, and positioned to challenge Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5. The release comes weeks after SpaceX formally exercised its option to acquire Cursor at a $60 billion valuation.

• Apple is reportedly lobbying the Trump administration for clearance to buy memory chips from blacklisted Chinese company CXMT, per the FT, as the memory shortage that drove last month’s Mac and iPad price hikes continues to bite; Apple wants guarantees CXMT won’t be added to the US Entity List (which would trigger licensing requirements) after CXMT’s Pentagon 1260H listing over alleged PLA ties.

• DeepSeek began developing its own AI inference chip to reduce reliance on Nvidia and Huawei after its first $7.4 billion fundraise at a $52-59 billion valuation, sending Nvidia shares down 1.6% on the news.

• Tesla expanded its Robotaxi service to Miami on July 3, its first city outside Texas and California, running fully unsupervised rides in a geofenced 10–14 square-mile stretch of western Miami-Dade from launch, raising Tesla’s active robotaxis to around 60 versus Waymo’s 4K vehicles and over 500K paid rides per week.

• Samsung forecast Q2 operating profit of 89.4 trillion won ($58.4 billion), a 19-fold year-on-year jump that overtakes Nvidia’s most recent quarterly profit and makes Samsung the world’s most profitable tech company for the period, driven almost entirely by AI memory demand.

At Contrary Research, we’ve built the best starting place to understand private tech companies. If you're interested in researching and writing about tech, apply here for our Research Fellowship!