Çelik here.

Two things shipped in 48 hours that change what an agent can do this week. OpenAI released a voice stack with GPT-5-class reasoning, live 70-language translation, and streaming transcription. AWS, Coinbase, and Stripe shipped the payments rail that lets agents transact in stablecoins with scoped, per-session wallets.

The same week, Anthropic doubled Claude Code's rate limits on the back of a new deal to take all of SpaceX's Colossus 1 data center, and an Albanian founding team in Vienna closed a $3.6M pre-seed for an AutoCAD challenger.

If you sell into customer service or any voice-heavy workflow, the OpenAI release is the reset for your stack. If you build agentic products, the AWS payments piece is the day the wallet showed up. If you operate in the Western Balkans, the Synaps round is the proof diaspora capital is starting to write checks. If you build agent products on a frontier API, the Anthropic-SpaceX compute deal tells you which lab just bought itself the next 18 months of capacity.

On May 7, OpenAI released three new audio models: GPT-Realtime-2 (voice with GPT-5-class reasoning, 128K context, adjustable reasoning levels), GPT-Realtime-Translate (live speech across 70+ input languages into 13 outputs), and GPT-Realtime-Whisper (streaming speech-to-text).

Realtime-2 (high) scored 15.2% above the prior model on Big Bench Audio. Pricing landed at $32 / $64 per million audio input/output tokens, $0.034/min for translation, $0.017/min for transcription.

The positional shift: voice agents move from demo to production this week. The combination of GPT-5 reasoning plus live translation plus streaming transcription is the stack a customer-service ops director has been waiting for. The model that answers a French speaker in Croatian while reading a CRM record in real time is now in a developer playground at consumer-grade prices.

If you run a contact center, a sales-call function, or any voice-heavy workflow, the move this week is to put GPT-Realtime-Translate against your top three call-language pairs and time the latency end to end. The vendors selling you per-language licenses are about to get a renegotiation. If you build a voice product, the new ceiling is set by Realtime-2's reasoning, not by ASR accuracy.

That’s 6 sales in 7 months. 29 all time. And the performance?

16.5%, 17.6%, and 17.8%, net annualized returns on sold works held longer than one year (See all 29 at Masterworks.com)

It’s not from stocks, private equity, or real estate… it’s from contemporary and post war art. Crazy, right?

With Masterworks, you don’t need to be a BILLIONAIRE to invest in multi-million dollar art anymore.

Historically, the segment overall has had attractive appreciation and low correlation to stocks.*

Masterworks targets works featuring legends like Banksy, Basquiat, and Picasso, identifying what they believe to have significant long-term appreciation potential, not just at the artist level but at the level of individual artworks.

As one of the largest players in the art market, with $1.3 billion invested over 500 artworks, they pass critical advantages through to their 70,000+ members to add art to their portfolios strategically.

Looking to diversify your investments in 2026?

Click here to skip the waitlist

^{*According to Masterworks data. Investing involves risk. Past performance is not indicative of future returns. See important Reg A disclosures at masterworks.com/cd.}

On May 7 AWS launched Amazon Bedrock AgentCore Payments with Coinbase and Stripe as the wallet rails. The first version focuses on micropayments: agents pay for APIs, data feeds, and paywalled content in USDC stablecoins, with future versions planned for hotel bookings and merchant transactions. Developers pick a Coinbase wallet or a Stripe Privy wallet; users explicitly authorize the agent and set per-session spending limits at runtime. The protocol is Coinbase's x402, an HTTP-native payment layer for agent-to-agent transactions.

The reframe is that the missing piece between "an agent can act" and "an agent can transact" just got installed by the largest cloud, the largest US crypto exchange, and the largest payments platform in a single coordinated launch. Stripe's Sessions 2026 reveal of the Link wallet for agents eight days ago set the pattern. AWS just turned it into a default Bedrock primitive. Agentic commerce stops being a 2027 phrase this week.

If you sell APIs, data products, or any per-call digital service, the move is to add x402-compatible billing to your endpoint before June. The first wave of agents that buy from each other will pay whoever shipped the integration first. If you build agent products, the user-experience question is whether the spending-limit UI is yours or AWS's, because that controls the trust relationship.

On May 5 Anthropic doubled Claude Code's five-hour rate limits across Pro, Max, Team, and seat-based Enterprise plans, removed peak-hours throttling for Pro and Max, and raised API rate limits across the Claude Opus family. The capacity comes from a new agreement to take the entire SpaceX Colossus 1 data center: more than 300 megawatts and over 220,000 NVIDIA GPUs online inside the month. Anthropic also said it is exploring multi-gigawatt orbital AI compute with SpaceX, with no contract signed yet.

The frame is the supply side of the same week's agent story. Voice agents and stablecoin-paying agents both burn tokens at production volumes nobody has run before, and the rate limits inside Claude Code, Cursor, and Bedrock are the ceiling on what an agent can do in a workday. Anthropic's Colossus 1 deal joins its existing stack: up to 5 GW with Amazon (about 1 GW online by end-2026), 5 GW with Google and Broadcom from 2027, $30B of Azure capacity with Microsoft and NVIDIA, and a $50B Fluidstack commitment. The number to watch is the gigawatt total, not the next model release.

If you build agent products on Claude, the May 5 changes mean your tightest production constraint just loosened. Re-test your worst-case Claude Code or API workflows against the new ceilings this week before reworking your pricing model. If you build on OpenAI or Google, expect parallel ceiling lifts within the quarter; the labs that don't will lose the agent workloads to the labs that will.

Five picks worth your attention.

Perplexity Personal Computer ships to every Mac Pro and Enterprise user. Available on the Mac desktop app from May 7, the agent reads and edits local files, drives native Mac apps, and runs the web through Comet, with a sandboxed audit trail and reversible actions. Perplexity recommends a Mac mini left running so the agent works in the background. If you live on a Mac, this is the productivity reset of the month.

Figma's /prototype-to-figma and figma-use-figjam MCP skills land. Figma's May release lets a coding agent take a working prototype and bring every screen onto the canvas as a design-system frame, plus read and write FigJam boards (stickies, sections, connectors, tables). Pair with generate-project-plan to convert a doc or codebase into a visual board. A real "AI moves into the design tool" moment for product teams.

Klaviyo Composer + expanded Canva integration generates campaigns from a prompt. Composer writes the subject line, copy, image, segment, and send logic in one go; the new Canva integration lets you design in Canva and pull layouts directly into Klaviyo for personalisation. If you run e-commerce or newsletter ops, this is the workflow that replaces three Tuesdays of templating.

Glean's Adaptive Reasoning + native connectors for Apollo, Dropbox, Granola. Glean's May launch adds an enterprise AI coworker that codifies repeatable workflows as Skills, delegates multiple workstreams across enterprise apps with approval controls, and auto-routes between models based on task complexity. The Apollo connector turns CRM-and-email research into a callable agent action.

Claude Code 2.1.133 ships /ultrareview, /less-permission-prompts, and xhigh effort. Anthropic's May 7 release adds /ultrareview for parallel multi-agent code review in the cloud, /less-permission-prompts which scans transcripts and proposes a prioritized allowlist for .claude/settings.json, and an xhigh effort level for Opus 4.7. The allowlist generator alone saves five clicks per session.That’s it for today