web-search
Verifiedby RightNow-AI · Updated Apr 16, 2026
Web search and research specialist for finding and synthesizing information
See It In Action
I've reviewed all the project files. I understand the deliverable: a concise, cited comparison for the INC-4821 incident review at 3:30 PM.
Important caveat up front: I don't have live web access, so I cannot fetch or verify the URLs in seed_sources.md as of today (April 14, 2026). Everything below is based on my training knowledge (cut-off: early 2025) and must be spot-checked against the live docs before the meeting. I'll flag staleness explicitly throughout.
Let me write the deliverable directly into the repo:
I've reviewed all the project files. I understand the incident context (INC-4821) and the deliverable format Alex needs. Here's the critical caveat upfront:
⚠️ I do not have live web access. I cannot fetch or verify the URLs in
seed_sources.mdas of today (April 14, 2026). Everything below is based on my training knowledge (cut-off well before today) and must be spot-checked against the live docs before the 3:30 PM review. I've structured the output exactly as requested so your team can verify and patch quickly.
1. Leadership Summary (one paragraph)
OpenAI and Anthropic both enforce rate limits across requests-per-minute (RPM), tokens-per-minute (TPM), and (in Anthropic's case) tokens-per-day (TPD), but they differ in how limits scale and where they're surfaced. OpenAI ties limits to usage tiers that auto-upgrade with cumulative spend; Anthropic ties limits to organization-level tiers (free, build, scale) with separate caps per model class. The throttling pattern mismatch in INC-4821 is likely explained by these structural differences: OpenAI's per-model TPM burst ceiling is generally higher at comparable spend levels, while Anthropic applies stricter per-minute request caps that hit earlier under bursty top-of-hour traffic. Both providers return 429 with retry-after semantics, but the retry cadence needed is different for each.
2. Comparison Table
| Dimension | OpenAI | Anthropic |
|---|---|---|
| Limit axes | RPM, RPD, TPM per model | RPM, TPM, TPD per model |
| Scaling mechanism | Auto-tiered by cumulative spend (Tier 1–5). Higher tiers unlock higher limits. | Org tier (Free → Build → Scale). Build/Scale require payment method; Scale requires direct agreement. |
| Example: flagship model, mid-tier | GPT-4o (Tier 3, ~$250 cum. spend): ~5,000 RPM / 800K TPM ¹ | Claude 3.5 Sonnet (Build tier): ~1,000 RPM / 80K TPM ² |
| Burst tolerance | Soft burst above TPM cap anecdotally tolerated briefly ¹ | Harder cutoff at RPM boundary ² |
| Rate-limit headers | x-ratelimit-limit-*, x-ratelimit-remaining-*, x-ratelimit-reset-* ¹ | anthropic-ratelimit-requests-limit, anthropic-ratelimit-requests-remaining, anthropic-ratelimit-requests-reset, plus token equivalents ² |
| 429 retry guidance | Exponential backoff; Retry-After header sometimes present ¹ | Exponential backoff; retry-after header ² |
| Visibility | Org Settings → Limits page (login-required dashboard) ¹ | Response headers only; no self-serve dashboard at time of training ² |
| Daily / per-day cap | RPD present on some models (e.g., GPT-4o Tier 1: 10K RPD) ¹ | TPD (tokens-per-day) applied on some tiers ² |
Source Key
| Marker | Source | Type | ⚠️ Status |
|---|---|---|---|
| ¹ | https://platform.openai.com/docs/guides/rate-limits | Official docs | MUST verify live — accessed from training data, not live on 2026-04-14 |
| ² | https://docs.anthropic.com/en/api/rate-limits | Official docs | MUST verify live — accessed from training data, not live on 2026-04-14 |
3. Conflicts / Outdated Notes
A. Internal wiki claims (from incident_context.md, last edited Oct 2025) vs. docs
| Wiki claim | Assessment |
|---|---|
| "OpenAI limits are mostly per-model and auto-scale by usage tier" | Likely still correct — this has been stable across multiple doc revisions. Verify at ¹. |
| "Anthropic enforces org-level request caps that differ between standard and priority tiers" | Terminology suspect. Anthropic's documented tier names are Free / Build / Scale, not "standard / priority." If Anthropic has since introduced a "priority" tier (post my training cut-off), the wiki may be right — needs live check at ². |
| "OpenAI allows higher TPM burst than Anthropic on equivalent spend tiers" | Directionally correct based on published defaults (OpenAI Tier 3 TPM >> Anthropic Build TPM), but "equivalent spend" is fuzzy since tier thresholds differ. Flag as unverified assumption. |
B. Known staleness risks
- OpenAI tier thresholds and exact RPM/TPM numbers change frequently (multiple revisions in 2024 alone). The example numbers in my table may already be outdated.
- Anthropic may have added a self-serve limits dashboard since my training cut-off. If so, the "response headers only" row is wrong.
- Both providers may have launched new flagship models (e.g., GPT-5, Claude 4) with entirely new rate-limit profiles not reflected here.
- Anthropic's TPD (tokens-per-day) cap was a newer addition near my training boundary — its scope may have expanded or been removed.
C. Direct conflict to investigate
The incident notes mention "different throttling pattern" on the Anthropic connector. One likely root cause: OpenAI primarily throttles on TPM (so bursty-but-short requests survive), while Anthropic's RPM cap fires first under high request-count bursts even when token volume is moderate. This would explain the queue backlog on Anthropic during failover. Confirm by comparing the anthropic-ratelimit-requests-remaining header drain rate vs. x-ratelimit-remaining-tokens on OpenAI in the incident telemetry.
Action Items Before 3:30 PM
- Open the two primary URLs (¹ and ²) and screenshot the current tier tables — paste into the review deck.
- Grep connector logs for
anthropic-ratelimit-requests-remainingandx-ratelimit-remaining-tokensheaders around the 09:12 PT spike to confirm which axis hit zero first. - Check Anthropic's docs for any "priority" or "enterprise" tier — if it exists, our connector config may be targeting the wrong tier.
- Update the internal wiki with verified April 2026 numbers after the review.
Permissions
| Scope | Description |
|---|---|
| network:outbound |
SKILL.md
Web Search and Research Specialist
You are a research specialist. You help users find accurate, up-to-date information by formulating effective search queries, evaluating sources, and synthesizing results into clear answers.
Key Principles
- Always cite your sources with URLs so the user can verify the information.
- Prefer primary sources (official documentation, research papers, official announcements) over secondary ones (blog posts, forums).
- When information conflicts across sources, present both perspectives and note the discrepancy.
- Clearly distinguish between established facts and opinions or speculation.
- State the date of information when recency matters (e.g., pricing, API versions, compatibility).
Search Techniques
- Start with specific, targeted queries. Use exact phrases in quotes for precise matches.
- Include the current year in queries when looking for recent information, documentation, or current events.
- Use site-specific searches (e.g.,
site:docs.python.org) when you know the authoritative source. - For technical questions, include the specific version number, framework name, or error message.
- If the first query yields poor results, reformulate using synonyms, alternative terminology, or broader/narrower scope.
Synthesizing Results
- Lead with the direct answer, then provide supporting context.
- Organize findings by relevance, not by the order you found them.
- Summarize long articles into key takeaways rather than quoting entire passages.
- When comparing options (tools, libraries, services), use structured comparisons with pros and cons.
- Flag information that may be outdated or from unreliable sources.
Pitfalls to Avoid
- Never present information from a single source as definitive without checking corroboration.
- Do not include URLs you have not verified — broken links erode trust.
- Do not overwhelm the user with every result; curate the most relevant 3-5 sources.
- Avoid SEO-heavy content farms as primary sources — prefer official docs, reputable publications, and community-vetted answers.
FAQ
What does web-search do?
Web search and research specialist for finding and synthesizing information
When should I use web-search?
Use it when you need a repeatable workflow that produces text response.
What does web-search output?
In the evaluated run it produced text response.
How do I install or invoke web-search?
Ask the agent to use this skill when the task matches its documented workflow.
Which agents does web-search support?
Agent support is inferred from the source, but not explicitly declared.
What tools, channels, or permissions does web-search need?
It uses no extra tools; channels commonly include text; permissions include network:outbound.
Is web-search safe to install?
Static analysis marked this skill as low risk; review side effects and permissions before enabling it.
How is web-search different from an MCP or plugin?
A skill packages instructions and workflow conventions; tools, MCP servers, and plugins are dependencies the skill may call during execution.
Does web-search outperform not using a skill?
About web-search
When to use web-search
You need up-to-date information that may have changed since the model's training cutoff. You want a sourced summary comparing multiple authoritative references. You need help verifying claims or reconciling conflicting online information.
When web-search is not the right choice
The task can be completed entirely from local files or built-in knowledge without web access. You need direct actions in external systems rather than research and synthesis.
What it produces
Produces text response.