Every Solana protocol that needs an external price ends up picking between two oracle networks: Pyth and Switchboard. The choice used to be obvious — Pyth for low-latency price feeds, Switchboard for everything else. In 2026 the lines have blurred, both networks have shipped major upgrades, and the right pick depends more on your specific use case than on broad reputation.
This is the comparison we wish someone had written when we were evaluating oracle integrations. Real numbers, real failure modes, real tradeoffs.
TL;DR table
| Dimension | Pyth | Switchboard |
|---|
| Push vs pull | Pull (most), push (some) | Pull (V3), push (V2 legacy) |
| Update latency | 400ms-1s typical | 1-5s typical |
| Asset coverage | 500+ feeds (crypto, equities, FX, commodities) | 300+ feeds, plus custom on-demand |
| Pricing model | Publishers stake PYTH, earn from fees | Operators stake SWITCH, earn from fees |
| Aggregation method | Aggregated medians + confidence intervals | Aggregated medians + standard deviations |
| Custom feeds | Limited (publishers control listings) | Yes — anyone can deploy a custom feed |
| EVM cross-chain | Yes (Pythnet bridge) | Yes (Switchboard On-Demand) |
| Used by | Drift, Kamino, Jupiter Perps, Flash Trade, MarginFi | Solend, Hubble, Streamflow, custom DeFi |
| Solana TVL using each | $4B+ | $800M+ |
| Confidence intervals | Yes (sigma + percent) | Yes (std deviation) |
If you're building a perpetuals protocol, lending market, or anything else where latency directly affects user economics, the baseline is Pyth. If you need custom data feeds — weather, sports, off-chain APIs, anything not in the standard asset list — Switchboard is the answer. We'll get into why below.
How each oracle actually works
Both networks share the same general architecture: a set of data publishers contribute prices, an on-chain aggregator combines them, smart contracts read the aggregated result. The differences are in how publishers are selected, how updates land on-chain, and what data types each supports.
Pyth. Data publishers are institutional trading firms — Jane Street, Jump Trading, Wintermute, Cumberland, and similar — that publish prices to Pyth's off-chain network (Pythnet, a Solana fork). The aggregator runs on Pythnet, produces a combined price with a confidence interval, and that result gets bridged to mainnet on demand.
The "pull" model is the key innovation. Instead of publishers pushing every update on-chain (which would burn enormous SOL on fees), the aggregated price sits on Pythnet and gets pulled into mainnet only when a transaction needs it. A user's swap transaction includes a Pyth update instruction that brings the latest price on-chain right before the swap executes. This makes each update effectively atomic with the transaction that uses it.
Latency: typical update freshness is 400ms to 1 second from publisher quote to on-chain availability. Confidence intervals scale with publisher disagreement — calm market = tight bounds, chaotic market = wide bounds.
Switchboard. Started with a different architecture: open participation. Anyone can run a Switchboard operator, anyone can deploy a custom feed by specifying the data source (a CoinGecko API, a sports scoreboard, an HTTP endpoint with a JSONPath). The operator network polls the source, signs the result, and aggregates on-chain.
Switchboard V3 (the current generation) introduced "on-demand" feeds — similar in spirit to Pyth's pull model. Instead of paying for continuous on-chain updates, integrators request a fresh fetch when needed, the operator network pulls in real-time and signs the result.
Latency: typical update freshness is 1-5 seconds. Slightly slower than Pyth because the operator network is more diverse and the consensus mechanism trades latency for robustness.
The other big Switchboard advantage: custom feeds. If you need the price of a stock that isn't in Pyth's list, or the score of a Premier League match, or the output of a specific API — Switchboard can do it, Pyth typically can't.
Performance under stress
Both oracles have been tested under real Solana network conditions. The interesting comparisons:
Solana network slowdowns. When Solana TPS drops (rare but historically real), oracles become a bottleneck. Pyth's pull model degrades more gracefully — the off-chain Pythnet keeps producing fresh prices, the only failure mode is that the on-chain pull instruction can't land. Switchboard's operator network similarly survives, but with more on-chain dependence the update gap can stretch longer during congestion.
In practice, both oracles handled the few real Solana incidents of 2025-2026 without major depegs or liquidation failures. Drift and Kamino (Pyth-based) and Solend (Switchboard-based) all maintained orderly operations during the slowdowns we saw.
Extreme volatility events. When BTC moves 10% in five minutes (e.g., the early-2026 ETF-flow flash), oracle confidence intervals widen dramatically. Both Pyth and Switchboard widened correctly during those events. The question is what your protocol does with the wider confidence interval.
Pyth's confidence intervals are aggressive: publishers can be slashed for prices outside the consensus, so they self-protect by widening sigma during chaos. Switchboard's standard deviations are similar but the operator economic incentives differ. Most lending protocols on Solana implement an "if confidence interval exceeds X% of price, pause liquidations" guard that works for both.
Malicious publisher scenarios. Pyth's slashing mechanic punishes publishers whose quote deviates from the consensus by more than a configurable percentage. This prevents one rogue publisher from moving the price, but introduces a different attack — coordinate enough publishers and you can move the consensus. The economic cost of doing so is real (millions of dollars at stake), but possible.
Switchboard's operator network is more diverse but individually less staked. The attack profile is wider but each individual oracle attack cheaper.
No serious oracle attack has succeeded on Solana to date. Both networks have held up.
