Writer Unveils Palmyra X6, Cutting AI Agent Costs by 52% as Token Spending Surges
Key Highlights
- The Geopolitical Foundation: Palmyra X6 is built on GLM-5.2, an open-weight mixture-of-experts model created by Beijing-based Z.ai (formerly Zhipu AI). Writer explicitly fine-tuned this open-source foundation on U.S. infrastructure.
- Surgical Fine-Tuning: The 744-billion-parameter model uses Anchored Supervised Fine-Tuning (ASFT) with just 626 curated synthetic trajectories, trained for a single epoch using the Muon optimizer to preserve base capabilities while teaching precise tool use.
- Competitive Pricing: Writer prices X6 at $2 per million input tokens / $8 per million output tokens—a fraction of Anthropic’s Claude Opus 4.8 ($15/$75).
The Token Cost Crisis
Standard Chatbot: Prompt ───────► Single Response (1x tokens)
Agentic AI Workflow: Prompt ───────► [Plan ↺ Tool Call ↺ Validate ↺ Retry] ───► Final Output (20x+ tokens)
"The biggest barrier today to enterprise expansion using AI is actually not model capabilities in most cases; it's actually the cost around them," explained Matan-Paul Shetrit, Director of Product Management at Writer. "Reducing the cost is not hurting my bottom line. It's actually expanding it, because it's expanding the TAM of opportunity within an organization."
Strategic Takeaways & Benchmark Performance
| Model | Average Score | Input / Output Price (per M Tokens) |
| Palmyra X6 | 0.87 | $2 / $8 |
| Claude Opus 4.8 | 0.86 | $15 / $75 |
| Claude Sonnet 4.6 | 0.85 | $3 / $15 |
| GPT-5.5 | 0.80 | Standard Enterprise |
| Gemini 3.1 | 0.77 | Standard Enterprise |
"It's very much a Palmyra model, and we just happen to grab the floating point numbers as the starting point... Public benchmarks let us know that we're climbing the right hill, but we don't slavishly follow them, because that's not really serving our customers."
Netflix has abruptly shuttered Night School and Moonloot, two of its gaming studios, the streamer told Variety Thursday. The news comes just weeks after Night School released its latest horror game, "Unhinged." Meanwhile, a yearslong divestment spree has narrowed the scope of Netflix's gaming business, which now includes just the Helsinki-based studio Next Games and a team working with outside developers. But Netflix isn't abandoning gaming initiatives altogether. Last week, the company said it would stream the upcoming extended trailer for Grand Theft Auto 6.
Google is shipping Gemini 3.7 Flash just three weeks after 3.6 — and it's aimed squarely at developers building coding agents and enterprise automation, with a temporary 50% API price cut to sweeten the deal.
Through the end of 2026, input tokens cost $0.75 per million and output $3.75 per million. On January 1, 2027, prices double. The pitch: evaluate now, lock in savings before the new year.
The model shows big gains in software engineering and workflow benchmarks. FrontierCode jumps from 34.4% to 43.6%, DeepSWE from 49% to 65.3%, and AutomationBench nearly doubles to 30.4%. Google says 3.7 Flash "thinks more diligently" — planning harder, recovering from errors, and completing multi-step tasks with fewer retries and less human babysitting.
But the results aren't universal. GPT-5.6 Terra still leads on several terminal and OS benchmarks, and Claude Sonnet 5 beats it on multimodal desktop tasks. Google's own numbers show a model that's cheaper and very competitive — not one that dominates everywhere.
The timing matters. Gemini 3.5 Pro remains MIA. Google's top general-purpose model is still Gemini 3.1 Pro from February, while DeepMind leadership has reshuffled and key researchers have left. Flash is getting faster and sharper; Pro is still stuck in partner testing.
Gemini 3.7 Flash isn't claiming the AI crown. It's making a different play — capable enough for serious work, cheap enough to run at scale, and available right now. Whether it's worth the full price in January depends on how reliably it ships real work, not how it scores on leaderboards.
A $2 trillion IPO is not a valuation. It is a vote.
The FT reports that Anthropic's investors expect an October listing north of $2 trillion, which would make it the largest public float in history, ahead of SpaceX. Management has not fixed a number. The company filed in June and is in a quiet period. What we have are investor models, not a prospectus.
So read it for what it is. Backers are underwriting a revenue run rate of $100 billion to $120 billion by year end against roughly $47 billion annualized in May. At 30 times revenue, the math gets you to $3 trillion. At 10 times, it gets you to $1.2 trillion. The entire gap between those two answers is one assumption: how long does the growth rate hold.
I spent enough years on the sell side to know that multiple expansion feels like insight and behaves like leverage.
Public markets will not price the story. They will price the second derivative.
What multiple would you underwrite here?
Once a brand like Unilever is signing 300,000 content creators to its ranks, the creator economy has expanded to the point that it's become "a labor market," advisor Alex Brownstein writes on LinkedIn. That, in turn, raises existential questions for creators: What happens when something that depends on feeling human becomes industrialized? Already, platforms are policing authenticity, Business Insider reports, increasingly labeling human-made content as AI-generated. As branded and synthetic content proliferates further, some creators may have to reckon with whether they're scaling themselves out of a job.
Google and SpaceXAI escalate AI price war with new discount models
SpaceXAI introduced Grok 4.6 on Wednesday, August 12, claiming its latest model matches frontier-level performance from OpenAI and Anthropic while undercutting them on price — the latest salvo in an intensifying AI price war that saw Google release its own discounted model just one day later.
Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens, making it roughly one-fourth the cost of Anthropic's Claude Opus 5 while delivering competitive results on key benchmarks. On the Artificial Analysis Intelligence Index, Grok 4.6 scored on par with OpenAI's GPT-5.6 Sol, according to SpaceXAI, though independent reviewers note it sits in third place behind Anthropic's Fable 5 and Opus 5. MarketWatch reported the model scored five points higher than its predecessor, Grok 4.5, on that index.
The model is built not as a new pre-training run but as a post-training and reinforcement learning upgrade on Grok 4.5, with a focus on long-running agents, multi-step reliability, and self-verification. SpaceXAI said Grok 4.6 can research topics, plan applications, build features, and refine results through multiple rounds of feedback. It is available immediately in Cursor and Grok Build.
An Escalating Price War
The launch comes amid a broader price collapse across frontier AI models. On August 13, Google released Gemini 3.7 Flash at $0.75 per million input tokens and $3.75 per million output tokens — half the original price of its predecessor, Gemini 3.6 Flash. OpenAI had recently cut the price of its GPT-5.6 Luna model by up to 80%. Chinese models from DeepSeek, Moonshot AI, and Alibaba have further pressured Western providers by offering near-frontier performance at a fraction of the cost.
Strengths and Limitations
Early independent testing suggests Grok 4.6 excels at serious coding and agentic work but falls short on design and visual generation tasks, where reviewers found its outputs lagged behind competitors. The model also uses more than 30 percent more tokens than Grok 4.5, making it slower and more expensive per task than its predecessor despite the competitive headline pricing. Elon Musk has already teased that Grok 4.7 is coming in three to four weeks.
AI firms recruit thousands of gig workers to train robots with wearable cameras
A new class of gig work is emerging across more than 50 countries, where thousands of contract workers strap cameras to their foreheads and don motion-tracking gloves to record the most ordinary of human tasks — folding laundry, washing dishes, assembling parts — so that humanoid robots can learn to replicate them.
The trend, detailed in a Bloomberg newsletter published Thursday, represents what may be the latest frontier of the gig economy: one where the product is not code, content moderation, or food delivery, but raw footage of human hands in motion.
A Growing Industry Built on Mundane Motion
Companies like Micro1, a Palo Alto-based robotics data firm, and Objectways, which operates collection centers in southern India, have built sprawling networks of workers who record themselves performing everyday physical tasks using wearable sensors and head-mounted cameras. Micro1 alone collects over 160,000 hours of video each month across thousands of unique environments in more than 75 countries, according to the company.
The footage feeds directly into training pipelines for humanoid robots being developed by Tesla, Figure AI, and Scale AI, among others. Robotics companies collectively spend more than $100 million annually on this type of real-world training data, a reflection of how difficult it remains to teach machines physical dexterity through simulation alone.
An MIT Technology Review investigation earlier this year profiled workers like Zeus, a medical student in Nigeria who earns $15 per hour filming himself making beds and cooking for Micro1. In India, the pay is far lower — roughly 250 rupees (about $2.60) per hour — where workers in regions like Tamil Nadu record household and factory tasks for Objectways.
Privacy and Labor Concerns
The arrangement raises questions that echo earlier waves of AI data labeling outsourced to developing economies. Workers often do not know which companies will ultimately use their footage, how long it will be retained, or whether it could be repurposed beyond robotics training. Because the workplace is frequently the worker's own home, cameras inevitably capture family members, personal spaces, and other unintended details.
The gig staffing platform Instawork has moved to formalize the practice, launching a wearable camera system called Instacore — five cameras mounted on the head, chest, and wrists connected to a compute backpack — designed for eight-hour shifts in commercial environments like kitchens and warehouses.
A Global Race for Embodied AI Data
Over $6 billion flowed into humanoid robotics in 2025, fueling demand for the kind of real-world manipulation data that no internet scrape can provide. China has pursued similar efforts through VR headsets and exoskeletons, with Beijing's Humanoid Robot Data Training Center operating on a $30 million budget across more than 10,000 square meters. The scramble underscores a basic bottleneck: before robots can work alongside humans, they first need to watch them.